OpenCloudOS-Kernel/lib/maple_tree.c

7178 lines
177 KiB
C
Raw Normal View History

Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
// SPDX-License-Identifier: GPL-2.0+
/*
* Maple Tree implementation
* Copyright (c) 2018-2022 Oracle Corporation
* Authors: Liam R. Howlett <Liam.Howlett@oracle.com>
* Matthew Wilcox <willy@infradead.org>
*/
/*
* DOC: Interesting implementation details of the Maple Tree
*
* Each node type has a number of slots for entries and a number of slots for
* pivots. In the case of dense nodes, the pivots are implied by the position
* and are simply the slot index + the minimum of the node.
*
* In regular B-Tree terms, pivots are called keys. The term pivot is used to
* indicate that the tree is specifying ranges, Pivots may appear in the
* subtree with an entry attached to the value where as keys are unique to a
* specific position of a B-tree. Pivot values are inclusive of the slot with
* the same index.
*
*
* The following illustrates the layout of a range64 nodes slots and pivots.
*
*
* Slots -> | 0 | 1 | 2 | ... | 12 | 13 | 14 | 15 |
*
* Implied maximum
* Pivot 14
* Pivot 13
* Pivot 12
* Pivot 11
* Pivot 2
* Pivot 1
* Pivot 0
* Implied minimum
*
* Slot contents:
* Internal (non-leaf) nodes contain pointers to other nodes.
* Leaf nodes contain entries.
*
* The location of interest is often referred to as an offset. All offsets have
* a slot, but the last offset has an implied pivot from the node above (or
* UINT_MAX for the root node.
*
* Ranges complicate certain write activities. When modifying any of
* the B-tree variants, it is known that one entry will either be added or
* deleted. When modifying the Maple Tree, one store operation may overwrite
* the entire data set, or one half of the tree, or the middle half of the tree.
*
*/
#include <linux/maple_tree.h>
#include <linux/xarray.h>
#include <linux/types.h>
#include <linux/export.h>
#include <linux/slab.h>
#include <linux/limits.h>
#include <asm/barrier.h>
#define CREATE_TRACE_POINTS
#include <trace/events/maple_tree.h>
#define MA_ROOT_PARENT 1
/*
* Maple state flags
* * MA_STATE_BULK - Bulk insert mode
* * MA_STATE_REBALANCE - Indicate a rebalance during bulk insert
* * MA_STATE_PREALLOC - Preallocated nodes, WARN_ON allocation
*/
#define MA_STATE_BULK 1
#define MA_STATE_REBALANCE 2
#define MA_STATE_PREALLOC 4
#define ma_parent_ptr(x) ((struct maple_pnode *)(x))
#define ma_mnode_ptr(x) ((struct maple_node *)(x))
#define ma_enode_ptr(x) ((struct maple_enode *)(x))
static struct kmem_cache *maple_node_cache;
#ifdef CONFIG_DEBUG_MAPLE_TREE
static const unsigned long mt_max[] = {
[maple_dense] = MAPLE_NODE_SLOTS,
[maple_leaf_64] = ULONG_MAX,
[maple_range_64] = ULONG_MAX,
[maple_arange_64] = ULONG_MAX,
};
#define mt_node_max(x) mt_max[mte_node_type(x)]
#endif
static const unsigned char mt_slots[] = {
[maple_dense] = MAPLE_NODE_SLOTS,
[maple_leaf_64] = MAPLE_RANGE64_SLOTS,
[maple_range_64] = MAPLE_RANGE64_SLOTS,
[maple_arange_64] = MAPLE_ARANGE64_SLOTS,
};
#define mt_slot_count(x) mt_slots[mte_node_type(x)]
static const unsigned char mt_pivots[] = {
[maple_dense] = 0,
[maple_leaf_64] = MAPLE_RANGE64_SLOTS - 1,
[maple_range_64] = MAPLE_RANGE64_SLOTS - 1,
[maple_arange_64] = MAPLE_ARANGE64_SLOTS - 1,
};
#define mt_pivot_count(x) mt_pivots[mte_node_type(x)]
static const unsigned char mt_min_slots[] = {
[maple_dense] = MAPLE_NODE_SLOTS / 2,
[maple_leaf_64] = (MAPLE_RANGE64_SLOTS / 2) - 2,
[maple_range_64] = (MAPLE_RANGE64_SLOTS / 2) - 2,
[maple_arange_64] = (MAPLE_ARANGE64_SLOTS / 2) - 1,
};
#define mt_min_slot_count(x) mt_min_slots[mte_node_type(x)]
#define MAPLE_BIG_NODE_SLOTS (MAPLE_RANGE64_SLOTS * 2 + 2)
#define MAPLE_BIG_NODE_GAPS (MAPLE_ARANGE64_SLOTS * 2 + 1)
struct maple_big_node {
struct maple_pnode *parent;
unsigned long pivot[MAPLE_BIG_NODE_SLOTS - 1];
union {
struct maple_enode *slot[MAPLE_BIG_NODE_SLOTS];
struct {
unsigned long padding[MAPLE_BIG_NODE_GAPS];
unsigned long gap[MAPLE_BIG_NODE_GAPS];
};
};
unsigned char b_end;
enum maple_type type;
};
/*
* The maple_subtree_state is used to build a tree to replace a segment of an
* existing tree in a more atomic way. Any walkers of the older tree will hit a
* dead node and restart on updates.
*/
struct maple_subtree_state {
struct ma_state *orig_l; /* Original left side of subtree */
struct ma_state *orig_r; /* Original right side of subtree */
struct ma_state *l; /* New left side of subtree */
struct ma_state *m; /* New middle of subtree (rare) */
struct ma_state *r; /* New right side of subtree */
struct ma_topiary *free; /* nodes to be freed */
struct ma_topiary *destroy; /* Nodes to be destroyed (walked and freed) */
struct maple_big_node *bn;
};
#ifdef CONFIG_KASAN_STACK
/* Prevent mas_wr_bnode() from exceeding the stack frame limit */
#define noinline_for_kasan noinline_for_stack
#else
#define noinline_for_kasan inline
#endif
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
/* Functions */
static inline struct maple_node *mt_alloc_one(gfp_t gfp)
{
return kmem_cache_alloc(maple_node_cache, gfp);
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
}
static inline int mt_alloc_bulk(gfp_t gfp, size_t size, void **nodes)
{
return kmem_cache_alloc_bulk(maple_node_cache, gfp, size, nodes);
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
}
static inline void mt_free_bulk(size_t size, void __rcu **nodes)
{
kmem_cache_free_bulk(maple_node_cache, size, (void **)nodes);
}
static void mt_free_rcu(struct rcu_head *head)
{
struct maple_node *node = container_of(head, struct maple_node, rcu);
kmem_cache_free(maple_node_cache, node);
}
/*
* ma_free_rcu() - Use rcu callback to free a maple node
* @node: The node to free
*
* The maple tree uses the parent pointer to indicate this node is no longer in
* use and will be freed.
*/
static void ma_free_rcu(struct maple_node *node)
{
node->parent = ma_parent_ptr(node);
call_rcu(&node->rcu, mt_free_rcu);
}
static void mas_set_height(struct ma_state *mas)
{
unsigned int new_flags = mas->tree->ma_flags;
new_flags &= ~MT_FLAGS_HEIGHT_MASK;
BUG_ON(mas->depth > MAPLE_HEIGHT_MAX);
new_flags |= mas->depth << MT_FLAGS_HEIGHT_OFFSET;
mas->tree->ma_flags = new_flags;
}
static unsigned int mas_mt_height(struct ma_state *mas)
{
return mt_height(mas->tree);
}
static inline enum maple_type mte_node_type(const struct maple_enode *entry)
{
return ((unsigned long)entry >> MAPLE_NODE_TYPE_SHIFT) &
MAPLE_NODE_TYPE_MASK;
}
static inline bool ma_is_dense(const enum maple_type type)
{
return type < maple_leaf_64;
}
static inline bool ma_is_leaf(const enum maple_type type)
{
return type < maple_range_64;
}
static inline bool mte_is_leaf(const struct maple_enode *entry)
{
return ma_is_leaf(mte_node_type(entry));
}
/*
* We also reserve values with the bottom two bits set to '10' which are
* below 4096
*/
static inline bool mt_is_reserved(const void *entry)
{
return ((unsigned long)entry < MAPLE_RESERVED_RANGE) &&
xa_is_internal(entry);
}
static inline void mas_set_err(struct ma_state *mas, long err)
{
mas->node = MA_ERROR(err);
}
static inline bool mas_is_ptr(struct ma_state *mas)
{
return mas->node == MAS_ROOT;
}
static inline bool mas_is_start(struct ma_state *mas)
{
return mas->node == MAS_START;
}
bool mas_is_err(struct ma_state *mas)
{
return xa_is_err(mas->node);
}
static inline bool mas_searchable(struct ma_state *mas)
{
if (mas_is_none(mas))
return false;
if (mas_is_ptr(mas))
return false;
return true;
}
static inline struct maple_node *mte_to_node(const struct maple_enode *entry)
{
return (struct maple_node *)((unsigned long)entry & ~MAPLE_NODE_MASK);
}
/*
* mte_to_mat() - Convert a maple encoded node to a maple topiary node.
* @entry: The maple encoded node
*
* Return: a maple topiary pointer
*/
static inline struct maple_topiary *mte_to_mat(const struct maple_enode *entry)
{
return (struct maple_topiary *)
((unsigned long)entry & ~MAPLE_NODE_MASK);
}
/*
* mas_mn() - Get the maple state node.
* @mas: The maple state
*
* Return: the maple node (not encoded - bare pointer).
*/
static inline struct maple_node *mas_mn(const struct ma_state *mas)
{
return mte_to_node(mas->node);
}
/*
* mte_set_node_dead() - Set a maple encoded node as dead.
* @mn: The maple encoded node.
*/
static inline void mte_set_node_dead(struct maple_enode *mn)
{
mte_to_node(mn)->parent = ma_parent_ptr(mte_to_node(mn));
smp_wmb(); /* Needed for RCU */
}
/* Bit 1 indicates the root is a node */
#define MAPLE_ROOT_NODE 0x02
/* maple_type stored bit 3-6 */
#define MAPLE_ENODE_TYPE_SHIFT 0x03
/* Bit 2 means a NULL somewhere below */
#define MAPLE_ENODE_NULL 0x04
static inline struct maple_enode *mt_mk_node(const struct maple_node *node,
enum maple_type type)
{
return (void *)((unsigned long)node |
(type << MAPLE_ENODE_TYPE_SHIFT) | MAPLE_ENODE_NULL);
}
static inline void *mte_mk_root(const struct maple_enode *node)
{
return (void *)((unsigned long)node | MAPLE_ROOT_NODE);
}
static inline void *mte_safe_root(const struct maple_enode *node)
{
return (void *)((unsigned long)node & ~MAPLE_ROOT_NODE);
}
static inline void *mte_set_full(const struct maple_enode *node)
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
{
return (void *)((unsigned long)node & ~MAPLE_ENODE_NULL);
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
}
static inline void *mte_clear_full(const struct maple_enode *node)
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
{
return (void *)((unsigned long)node | MAPLE_ENODE_NULL);
}
static inline bool mte_has_null(const struct maple_enode *node)
{
return (unsigned long)node & MAPLE_ENODE_NULL;
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
}
static inline bool ma_is_root(struct maple_node *node)
{
return ((unsigned long)node->parent & MA_ROOT_PARENT);
}
static inline bool mte_is_root(const struct maple_enode *node)
{
return ma_is_root(mte_to_node(node));
}
static inline bool mas_is_root_limits(const struct ma_state *mas)
{
return !mas->min && mas->max == ULONG_MAX;
}
static inline bool mt_is_alloc(struct maple_tree *mt)
{
return (mt->ma_flags & MT_FLAGS_ALLOC_RANGE);
}
/*
* The Parent Pointer
* Excluding root, the parent pointer is 256B aligned like all other tree nodes.
* When storing a 32 or 64 bit values, the offset can fit into 5 bits. The 16
* bit values need an extra bit to store the offset. This extra bit comes from
* a reuse of the last bit in the node type. This is possible by using bit 1 to
* indicate if bit 2 is part of the type or the slot.
*
* Note types:
* 0x??1 = Root
* 0x?00 = 16 bit nodes
* 0x010 = 32 bit nodes
* 0x110 = 64 bit nodes
*
* Slot size and alignment
* 0b??1 : Root
* 0b?00 : 16 bit values, type in 0-1, slot in 2-7
* 0b010 : 32 bit values, type in 0-2, slot in 3-7
* 0b110 : 64 bit values, type in 0-2, slot in 3-7
*/
#define MAPLE_PARENT_ROOT 0x01
#define MAPLE_PARENT_SLOT_SHIFT 0x03
#define MAPLE_PARENT_SLOT_MASK 0xF8
#define MAPLE_PARENT_16B_SLOT_SHIFT 0x02
#define MAPLE_PARENT_16B_SLOT_MASK 0xFC
#define MAPLE_PARENT_RANGE64 0x06
#define MAPLE_PARENT_RANGE32 0x04
#define MAPLE_PARENT_NOT_RANGE16 0x02
/*
* mte_parent_shift() - Get the parent shift for the slot storage.
* @parent: The parent pointer cast as an unsigned long
* Return: The shift into that pointer to the star to of the slot
*/
static inline unsigned long mte_parent_shift(unsigned long parent)
{
/* Note bit 1 == 0 means 16B */
if (likely(parent & MAPLE_PARENT_NOT_RANGE16))
return MAPLE_PARENT_SLOT_SHIFT;
return MAPLE_PARENT_16B_SLOT_SHIFT;
}
/*
* mte_parent_slot_mask() - Get the slot mask for the parent.
* @parent: The parent pointer cast as an unsigned long.
* Return: The slot mask for that parent.
*/
static inline unsigned long mte_parent_slot_mask(unsigned long parent)
{
/* Note bit 1 == 0 means 16B */
if (likely(parent & MAPLE_PARENT_NOT_RANGE16))
return MAPLE_PARENT_SLOT_MASK;
return MAPLE_PARENT_16B_SLOT_MASK;
}
/*
* mas_parent_enum() - Return the maple_type of the parent from the stored
* parent type.
* @mas: The maple state
* @node: The maple_enode to extract the parent's enum
* Return: The node->parent maple_type
*/
static inline
enum maple_type mte_parent_enum(struct maple_enode *p_enode,
struct maple_tree *mt)
{
unsigned long p_type;
p_type = (unsigned long)p_enode;
if (p_type & MAPLE_PARENT_ROOT)
return 0; /* Validated in the caller. */
p_type &= MAPLE_NODE_MASK;
p_type = p_type & ~(MAPLE_PARENT_ROOT | mte_parent_slot_mask(p_type));
switch (p_type) {
case MAPLE_PARENT_RANGE64: /* or MAPLE_PARENT_ARANGE64 */
if (mt_is_alloc(mt))
return maple_arange_64;
return maple_range_64;
}
return 0;
}
static inline
enum maple_type mas_parent_enum(struct ma_state *mas, struct maple_enode *enode)
{
return mte_parent_enum(ma_enode_ptr(mte_to_node(enode)->parent), mas->tree);
}
/*
* mte_set_parent() - Set the parent node and encode the slot
* @enode: The encoded maple node.
* @parent: The encoded maple node that is the parent of @enode.
* @slot: The slot that @enode resides in @parent.
*
* Slot number is encoded in the enode->parent bit 3-6 or 2-6, depending on the
* parent type.
*/
static inline
void mte_set_parent(struct maple_enode *enode, const struct maple_enode *parent,
unsigned char slot)
{
unsigned long val = (unsigned long)parent;
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
unsigned long shift;
unsigned long type;
enum maple_type p_type = mte_node_type(parent);
BUG_ON(p_type == maple_dense);
BUG_ON(p_type == maple_leaf_64);
switch (p_type) {
case maple_range_64:
case maple_arange_64:
shift = MAPLE_PARENT_SLOT_SHIFT;
type = MAPLE_PARENT_RANGE64;
break;
default:
case maple_dense:
case maple_leaf_64:
shift = type = 0;
break;
}
val &= ~MAPLE_NODE_MASK; /* Clear all node metadata in parent */
val |= (slot << shift) | type;
mte_to_node(enode)->parent = ma_parent_ptr(val);
}
/*
* mte_parent_slot() - get the parent slot of @enode.
* @enode: The encoded maple node.
*
* Return: The slot in the parent node where @enode resides.
*/
static inline unsigned int mte_parent_slot(const struct maple_enode *enode)
{
unsigned long val = (unsigned long)mte_to_node(enode)->parent;
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
if (val & MA_ROOT_PARENT)
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
return 0;
/*
* Okay to use MAPLE_PARENT_16B_SLOT_MASK as the last bit will be lost
* by shift if the parent shift is MAPLE_PARENT_SLOT_SHIFT
*/
return (val & MAPLE_PARENT_16B_SLOT_MASK) >> mte_parent_shift(val);
}
/*
* mte_parent() - Get the parent of @node.
* @node: The encoded maple node.
*
* Return: The parent maple node.
*/
static inline struct maple_node *mte_parent(const struct maple_enode *enode)
{
return (void *)((unsigned long)
(mte_to_node(enode)->parent) & ~MAPLE_NODE_MASK);
}
/*
* ma_dead_node() - check if the @enode is dead.
* @enode: The encoded maple node
*
* Return: true if dead, false otherwise.
*/
static inline bool ma_dead_node(const struct maple_node *node)
{
struct maple_node *parent = (void *)((unsigned long)
node->parent & ~MAPLE_NODE_MASK);
return (parent == node);
}
/*
* mte_dead_node() - check if the @enode is dead.
* @enode: The encoded maple node
*
* Return: true if dead, false otherwise.
*/
static inline bool mte_dead_node(const struct maple_enode *enode)
{
struct maple_node *parent, *node;
node = mte_to_node(enode);
parent = mte_parent(enode);
return (parent == node);
}
/*
* mas_allocated() - Get the number of nodes allocated in a maple state.
* @mas: The maple state
*
* The ma_state alloc member is overloaded to hold a pointer to the first
* allocated node or to the number of requested nodes to allocate. If bit 0 is
* set, then the alloc contains the number of requested nodes. If there is an
* allocated node, then the total allocated nodes is in that node.
*
* Return: The total number of nodes allocated
*/
static inline unsigned long mas_allocated(const struct ma_state *mas)
{
if (!mas->alloc || ((unsigned long)mas->alloc & 0x1))
return 0;
return mas->alloc->total;
}
/*
* mas_set_alloc_req() - Set the requested number of allocations.
* @mas: the maple state
* @count: the number of allocations.
*
* The requested number of allocations is either in the first allocated node,
* located in @mas->alloc->request_count, or directly in @mas->alloc if there is
* no allocated node. Set the request either in the node or do the necessary
* encoding to store in @mas->alloc directly.
*/
static inline void mas_set_alloc_req(struct ma_state *mas, unsigned long count)
{
if (!mas->alloc || ((unsigned long)mas->alloc & 0x1)) {
if (!count)
mas->alloc = NULL;
else
mas->alloc = (struct maple_alloc *)(((count) << 1U) | 1U);
return;
}
mas->alloc->request_count = count;
}
/*
* mas_alloc_req() - get the requested number of allocations.
* @mas: The maple state
*
* The alloc count is either stored directly in @mas, or in
* @mas->alloc->request_count if there is at least one node allocated. Decode
* the request count if it's stored directly in @mas->alloc.
*
* Return: The allocation request count.
*/
static inline unsigned int mas_alloc_req(const struct ma_state *mas)
{
if ((unsigned long)mas->alloc & 0x1)
return (unsigned long)(mas->alloc) >> 1;
else if (mas->alloc)
return mas->alloc->request_count;
return 0;
}
/*
* ma_pivots() - Get a pointer to the maple node pivots.
* @node - the maple node
* @type - the node type
*
* Return: A pointer to the maple node pivots
*/
static inline unsigned long *ma_pivots(struct maple_node *node,
enum maple_type type)
{
switch (type) {
case maple_arange_64:
return node->ma64.pivot;
case maple_range_64:
case maple_leaf_64:
return node->mr64.pivot;
case maple_dense:
return NULL;
}
return NULL;
}
/*
* ma_gaps() - Get a pointer to the maple node gaps.
* @node - the maple node
* @type - the node type
*
* Return: A pointer to the maple node gaps
*/
static inline unsigned long *ma_gaps(struct maple_node *node,
enum maple_type type)
{
switch (type) {
case maple_arange_64:
return node->ma64.gap;
case maple_range_64:
case maple_leaf_64:
case maple_dense:
return NULL;
}
return NULL;
}
/*
* mte_pivot() - Get the pivot at @piv of the maple encoded node.
* @mn: The maple encoded node.
* @piv: The pivot.
*
* Return: the pivot at @piv of @mn.
*/
static inline unsigned long mte_pivot(const struct maple_enode *mn,
unsigned char piv)
{
struct maple_node *node = mte_to_node(mn);
enum maple_type type = mte_node_type(mn);
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
if (piv >= mt_pivots[type]) {
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
WARN_ON(1);
return 0;
}
switch (type) {
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
case maple_arange_64:
return node->ma64.pivot[piv];
case maple_range_64:
case maple_leaf_64:
return node->mr64.pivot[piv];
case maple_dense:
return 0;
}
return 0;
}
/*
* mas_safe_pivot() - get the pivot at @piv or mas->max.
* @mas: The maple state
* @pivots: The pointer to the maple node pivots
* @piv: The pivot to fetch
* @type: The maple node type
*
* Return: The pivot at @piv within the limit of the @pivots array, @mas->max
* otherwise.
*/
static inline unsigned long
mas_safe_pivot(const struct ma_state *mas, unsigned long *pivots,
unsigned char piv, enum maple_type type)
{
if (piv >= mt_pivots[type])
return mas->max;
return pivots[piv];
}
/*
* mas_safe_min() - Return the minimum for a given offset.
* @mas: The maple state
* @pivots: The pointer to the maple node pivots
* @offset: The offset into the pivot array
*
* Return: The minimum range value that is contained in @offset.
*/
static inline unsigned long
mas_safe_min(struct ma_state *mas, unsigned long *pivots, unsigned char offset)
{
if (likely(offset))
return pivots[offset - 1] + 1;
return mas->min;
}
/*
* mas_logical_pivot() - Get the logical pivot of a given offset.
* @mas: The maple state
* @pivots: The pointer to the maple node pivots
* @offset: The offset into the pivot array
* @type: The maple node type
*
* When there is no value at a pivot (beyond the end of the data), then the
* pivot is actually @mas->max.
*
* Return: the logical pivot of a given @offset.
*/
static inline unsigned long
mas_logical_pivot(struct ma_state *mas, unsigned long *pivots,
unsigned char offset, enum maple_type type)
{
unsigned long lpiv = mas_safe_pivot(mas, pivots, offset, type);
if (likely(lpiv))
return lpiv;
if (likely(offset))
return mas->max;
return lpiv;
}
/*
* mte_set_pivot() - Set a pivot to a value in an encoded maple node.
* @mn: The encoded maple node
* @piv: The pivot offset
* @val: The value of the pivot
*/
static inline void mte_set_pivot(struct maple_enode *mn, unsigned char piv,
unsigned long val)
{
struct maple_node *node = mte_to_node(mn);
enum maple_type type = mte_node_type(mn);
BUG_ON(piv >= mt_pivots[type]);
switch (type) {
default:
case maple_range_64:
case maple_leaf_64:
node->mr64.pivot[piv] = val;
break;
case maple_arange_64:
node->ma64.pivot[piv] = val;
break;
case maple_dense:
break;
}
}
/*
* ma_slots() - Get a pointer to the maple node slots.
* @mn: The maple node
* @mt: The maple node type
*
* Return: A pointer to the maple node slots
*/
static inline void __rcu **ma_slots(struct maple_node *mn, enum maple_type mt)
{
switch (mt) {
default:
case maple_arange_64:
return mn->ma64.slot;
case maple_range_64:
case maple_leaf_64:
return mn->mr64.slot;
case maple_dense:
return mn->slot;
}
}
static inline bool mt_locked(const struct maple_tree *mt)
{
return mt_external_lock(mt) ? mt_lock_is_held(mt) :
lockdep_is_held(&mt->ma_lock);
}
static inline void *mt_slot(const struct maple_tree *mt,
void __rcu **slots, unsigned char offset)
{
return rcu_dereference_check(slots[offset], mt_locked(mt));
}
/*
* mas_slot_locked() - Get the slot value when holding the maple tree lock.
* @mas: The maple state
* @slots: The pointer to the slots
* @offset: The offset into the slots array to fetch
*
* Return: The entry stored in @slots at the @offset.
*/
static inline void *mas_slot_locked(struct ma_state *mas, void __rcu **slots,
unsigned char offset)
{
return rcu_dereference_protected(slots[offset], mt_locked(mas->tree));
}
/*
* mas_slot() - Get the slot value when not holding the maple tree lock.
* @mas: The maple state
* @slots: The pointer to the slots
* @offset: The offset into the slots array to fetch
*
* Return: The entry stored in @slots at the @offset
*/
static inline void *mas_slot(struct ma_state *mas, void __rcu **slots,
unsigned char offset)
{
return mt_slot(mas->tree, slots, offset);
}
/*
* mas_root() - Get the maple tree root.
* @mas: The maple state.
*
* Return: The pointer to the root of the tree
*/
static inline void *mas_root(struct ma_state *mas)
{
return rcu_dereference_check(mas->tree->ma_root, mt_locked(mas->tree));
}
static inline void *mt_root_locked(struct maple_tree *mt)
{
return rcu_dereference_protected(mt->ma_root, mt_locked(mt));
}
/*
* mas_root_locked() - Get the maple tree root when holding the maple tree lock.
* @mas: The maple state.
*
* Return: The pointer to the root of the tree
*/
static inline void *mas_root_locked(struct ma_state *mas)
{
return mt_root_locked(mas->tree);
}
static inline struct maple_metadata *ma_meta(struct maple_node *mn,
enum maple_type mt)
{
switch (mt) {
case maple_arange_64:
return &mn->ma64.meta;
default:
return &mn->mr64.meta;
}
}
/*
* ma_set_meta() - Set the metadata information of a node.
* @mn: The maple node
* @mt: The maple node type
* @offset: The offset of the highest sub-gap in this node.
* @end: The end of the data in this node.
*/
static inline void ma_set_meta(struct maple_node *mn, enum maple_type mt,
unsigned char offset, unsigned char end)
{
struct maple_metadata *meta = ma_meta(mn, mt);
meta->gap = offset;
meta->end = end;
}
/*
* ma_meta_end() - Get the data end of a node from the metadata
* @mn: The maple node
* @mt: The maple node type
*/
static inline unsigned char ma_meta_end(struct maple_node *mn,
enum maple_type mt)
{
struct maple_metadata *meta = ma_meta(mn, mt);
return meta->end;
}
/*
* ma_meta_gap() - Get the largest gap location of a node from the metadata
* @mn: The maple node
* @mt: The maple node type
*/
static inline unsigned char ma_meta_gap(struct maple_node *mn,
enum maple_type mt)
{
BUG_ON(mt != maple_arange_64);
return mn->ma64.meta.gap;
}
/*
* ma_set_meta_gap() - Set the largest gap location in a nodes metadata
* @mn: The maple node
* @mn: The maple node type
* @offset: The location of the largest gap.
*/
static inline void ma_set_meta_gap(struct maple_node *mn, enum maple_type mt,
unsigned char offset)
{
struct maple_metadata *meta = ma_meta(mn, mt);
meta->gap = offset;
}
/*
* mat_add() - Add a @dead_enode to the ma_topiary of a list of dead nodes.
* @mat - the ma_topiary, a linked list of dead nodes.
* @dead_enode - the node to be marked as dead and added to the tail of the list
*
* Add the @dead_enode to the linked list in @mat.
*/
static inline void mat_add(struct ma_topiary *mat,
struct maple_enode *dead_enode)
{
mte_set_node_dead(dead_enode);
mte_to_mat(dead_enode)->next = NULL;
if (!mat->tail) {
mat->tail = mat->head = dead_enode;
return;
}
mte_to_mat(mat->tail)->next = dead_enode;
mat->tail = dead_enode;
}
static void mte_destroy_walk(struct maple_enode *, struct maple_tree *);
static inline void mas_free(struct ma_state *mas, struct maple_enode *used);
/*
* mas_mat_free() - Free all nodes in a dead list.
* @mas - the maple state
* @mat - the ma_topiary linked list of dead nodes to free.
*
* Free walk a dead list.
*/
static void mas_mat_free(struct ma_state *mas, struct ma_topiary *mat)
{
struct maple_enode *next;
while (mat->head) {
next = mte_to_mat(mat->head)->next;
mas_free(mas, mat->head);
mat->head = next;
}
}
/*
* mas_mat_destroy() - Free all nodes and subtrees in a dead list.
* @mas - the maple state
* @mat - the ma_topiary linked list of dead nodes to free.
*
* Destroy walk a dead list.
*/
static void mas_mat_destroy(struct ma_state *mas, struct ma_topiary *mat)
{
struct maple_enode *next;
while (mat->head) {
next = mte_to_mat(mat->head)->next;
mte_destroy_walk(mat->head, mat->mtree);
mat->head = next;
}
}
/*
* mas_descend() - Descend into the slot stored in the ma_state.
* @mas - the maple state.
*
* Note: Not RCU safe, only use in write side or debug code.
*/
static inline void mas_descend(struct ma_state *mas)
{
enum maple_type type;
unsigned long *pivots;
struct maple_node *node;
void __rcu **slots;
node = mas_mn(mas);
type = mte_node_type(mas->node);
pivots = ma_pivots(node, type);
slots = ma_slots(node, type);
if (mas->offset)
mas->min = pivots[mas->offset - 1] + 1;
mas->max = mas_safe_pivot(mas, pivots, mas->offset, type);
mas->node = mas_slot(mas, slots, mas->offset);
}
/*
* mte_set_gap() - Set a maple node gap.
* @mn: The encoded maple node
* @gap: The offset of the gap to set
* @val: The gap value
*/
static inline void mte_set_gap(const struct maple_enode *mn,
unsigned char gap, unsigned long val)
{
switch (mte_node_type(mn)) {
default:
break;
case maple_arange_64:
mte_to_node(mn)->ma64.gap[gap] = val;
break;
}
}
/*
* mas_ascend() - Walk up a level of the tree.
* @mas: The maple state
*
* Sets the @mas->max and @mas->min to the correct values when walking up. This
* may cause several levels of walking up to find the correct min and max.
* May find a dead node which will cause a premature return.
* Return: 1 on dead node, 0 otherwise
*/
static int mas_ascend(struct ma_state *mas)
{
struct maple_enode *p_enode; /* parent enode. */
struct maple_enode *a_enode; /* ancestor enode. */
struct maple_node *a_node; /* ancestor node. */
struct maple_node *p_node; /* parent node. */
unsigned char a_slot;
enum maple_type a_type;
unsigned long min, max;
unsigned long *pivots;
unsigned char offset;
bool set_max = false, set_min = false;
a_node = mas_mn(mas);
if (ma_is_root(a_node)) {
mas->offset = 0;
return 0;
}
p_node = mte_parent(mas->node);
if (unlikely(a_node == p_node))
return 1;
a_type = mas_parent_enum(mas, mas->node);
offset = mte_parent_slot(mas->node);
a_enode = mt_mk_node(p_node, a_type);
/* Check to make sure all parent information is still accurate */
if (p_node != mte_parent(mas->node))
return 1;
mas->node = a_enode;
mas->offset = offset;
if (mte_is_root(a_enode)) {
mas->max = ULONG_MAX;
mas->min = 0;
return 0;
}
min = 0;
max = ULONG_MAX;
do {
p_enode = a_enode;
a_type = mas_parent_enum(mas, p_enode);
a_node = mte_parent(p_enode);
a_slot = mte_parent_slot(p_enode);
pivots = ma_pivots(a_node, a_type);
a_enode = mt_mk_node(a_node, a_type);
if (!set_min && a_slot) {
set_min = true;
min = pivots[a_slot - 1] + 1;
}
if (!set_max && a_slot < mt_pivots[a_type]) {
set_max = true;
max = pivots[a_slot];
}
if (unlikely(ma_dead_node(a_node)))
return 1;
if (unlikely(ma_is_root(a_node)))
break;
} while (!set_min || !set_max);
mas->max = max;
mas->min = min;
return 0;
}
/*
* mas_pop_node() - Get a previously allocated maple node from the maple state.
* @mas: The maple state
*
* Return: A pointer to a maple node.
*/
static inline struct maple_node *mas_pop_node(struct ma_state *mas)
{
struct maple_alloc *ret, *node = mas->alloc;
unsigned long total = mas_allocated(mas);
unsigned int req = mas_alloc_req(mas);
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
/* nothing or a request pending. */
if (WARN_ON(!total))
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
return NULL;
if (total == 1) {
/* single allocation in this ma_state */
mas->alloc = NULL;
ret = node;
goto single_node;
}
if (node->node_count == 1) {
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
/* Single allocation in this node. */
mas->alloc = node->slot[0];
mas->alloc->total = node->total - 1;
ret = node;
goto new_head;
}
node->total--;
ret = node->slot[--node->node_count];
node->slot[node->node_count] = NULL;
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
single_node:
new_head:
if (req) {
req++;
mas_set_alloc_req(mas, req);
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
}
memset(ret, 0, sizeof(*ret));
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
return (struct maple_node *)ret;
}
/*
* mas_push_node() - Push a node back on the maple state allocation.
* @mas: The maple state
* @used: The used maple node
*
* Stores the maple node back into @mas->alloc for reuse. Updates allocated and
* requested node count as necessary.
*/
static inline void mas_push_node(struct ma_state *mas, struct maple_node *used)
{
struct maple_alloc *reuse = (struct maple_alloc *)used;
struct maple_alloc *head = mas->alloc;
unsigned long count;
unsigned int requested = mas_alloc_req(mas);
count = mas_allocated(mas);
reuse->request_count = 0;
reuse->node_count = 0;
if (count && (head->node_count < MAPLE_ALLOC_SLOTS)) {
head->slot[head->node_count++] = reuse;
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
head->total++;
goto done;
}
reuse->total = 1;
if ((head) && !((unsigned long)head & 0x1)) {
reuse->slot[0] = head;
reuse->node_count = 1;
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
reuse->total += head->total;
}
mas->alloc = reuse;
done:
if (requested > 1)
mas_set_alloc_req(mas, requested - 1);
}
/*
* mas_alloc_nodes() - Allocate nodes into a maple state
* @mas: The maple state
* @gfp: The GFP Flags
*/
static inline void mas_alloc_nodes(struct ma_state *mas, gfp_t gfp)
{
struct maple_alloc *node;
unsigned long allocated = mas_allocated(mas);
unsigned int requested = mas_alloc_req(mas);
unsigned int count;
void **slots = NULL;
unsigned int max_req = 0;
if (!requested)
return;
mas_set_alloc_req(mas, 0);
if (mas->mas_flags & MA_STATE_PREALLOC) {
if (allocated)
return;
WARN_ON(!allocated);
}
if (!allocated || mas->alloc->node_count == MAPLE_ALLOC_SLOTS) {
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
node = (struct maple_alloc *)mt_alloc_one(gfp);
if (!node)
goto nomem_one;
if (allocated) {
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
node->slot[0] = mas->alloc;
node->node_count = 1;
} else {
node->node_count = 0;
}
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
mas->alloc = node;
node->total = ++allocated;
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
requested--;
}
node = mas->alloc;
node->request_count = 0;
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
while (requested) {
max_req = MAPLE_ALLOC_SLOTS;
if (node->node_count) {
unsigned int offset = node->node_count;
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
slots = (void **)&node->slot[offset];
max_req -= offset;
} else {
slots = (void **)&node->slot;
}
max_req = min(requested, max_req);
count = mt_alloc_bulk(gfp, max_req, slots);
if (!count)
goto nomem_bulk;
node->node_count += count;
allocated += count;
node = node->slot[0];
node->node_count = 0;
node->request_count = 0;
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
requested -= count;
}
mas->alloc->total = allocated;
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
return;
nomem_bulk:
/* Clean up potential freed allocations on bulk failure */
memset(slots, 0, max_req * sizeof(unsigned long));
nomem_one:
mas_set_alloc_req(mas, requested);
if (mas->alloc && !(((unsigned long)mas->alloc & 0x1)))
mas->alloc->total = allocated;
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
mas_set_err(mas, -ENOMEM);
}
/*
* mas_free() - Free an encoded maple node
* @mas: The maple state
* @used: The encoded maple node to free.
*
* Uses rcu free if necessary, pushes @used back on the maple state allocations
* otherwise.
*/
static inline void mas_free(struct ma_state *mas, struct maple_enode *used)
{
struct maple_node *tmp = mte_to_node(used);
if (mt_in_rcu(mas->tree))
ma_free_rcu(tmp);
else
mas_push_node(mas, tmp);
}
/*
* mas_node_count() - Check if enough nodes are allocated and request more if
* there is not enough nodes.
* @mas: The maple state
* @count: The number of nodes needed
* @gfp: the gfp flags
*/
static void mas_node_count_gfp(struct ma_state *mas, int count, gfp_t gfp)
{
unsigned long allocated = mas_allocated(mas);
if (allocated < count) {
mas_set_alloc_req(mas, count - allocated);
mas_alloc_nodes(mas, gfp);
}
}
/*
* mas_node_count() - Check if enough nodes are allocated and request more if
* there is not enough nodes.
* @mas: The maple state
* @count: The number of nodes needed
*
* Note: Uses GFP_NOWAIT | __GFP_NOWARN for gfp flags.
*/
static void mas_node_count(struct ma_state *mas, int count)
{
return mas_node_count_gfp(mas, count, GFP_NOWAIT | __GFP_NOWARN);
}
/*
* mas_start() - Sets up maple state for operations.
* @mas: The maple state.
*
* If mas->node == MAS_START, then set the min, max and depth to
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
* defaults.
*
* Return:
* - If mas->node is an error or not MAS_START, return NULL.
* - If it's an empty tree: NULL & mas->node == MAS_NONE
* - If it's a single entry: The entry & mas->node == MAS_ROOT
* - If it's a tree: NULL & mas->node == safe root node.
*/
static inline struct maple_enode *mas_start(struct ma_state *mas)
{
if (likely(mas_is_start(mas))) {
struct maple_enode *root;
mas->min = 0;
mas->max = ULONG_MAX;
mas->depth = 0;
root = mas_root(mas);
/* Tree with nodes */
if (likely(xa_is_node(root))) {
mas->depth = 1;
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
mas->node = mte_safe_root(root);
mas->offset = 0;
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
return NULL;
}
/* empty tree */
if (unlikely(!root)) {
mas->node = MAS_NONE;
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
mas->offset = MAPLE_NODE_SLOTS;
return NULL;
}
/* Single entry tree */
mas->node = MAS_ROOT;
mas->offset = MAPLE_NODE_SLOTS;
/* Single entry tree. */
if (mas->index > 0)
return NULL;
return root;
}
return NULL;
}
/*
* ma_data_end() - Find the end of the data in a node.
* @node: The maple node
* @type: The maple node type
* @pivots: The array of pivots in the node
* @max: The maximum value in the node
*
* Uses metadata to find the end of the data when possible.
* Return: The zero indexed last slot with data (may be null).
*/
static inline unsigned char ma_data_end(struct maple_node *node,
enum maple_type type,
unsigned long *pivots,
unsigned long max)
{
unsigned char offset;
if (type == maple_arange_64)
return ma_meta_end(node, type);
offset = mt_pivots[type] - 1;
if (likely(!pivots[offset]))
return ma_meta_end(node, type);
if (likely(pivots[offset] == max))
return offset;
return mt_pivots[type];
}
/*
* mas_data_end() - Find the end of the data (slot).
* @mas: the maple state
*
* This method is optimized to check the metadata of a node if the node type
* supports data end metadata.
*
* Return: The zero indexed last slot with data (may be null).
*/
static inline unsigned char mas_data_end(struct ma_state *mas)
{
enum maple_type type;
struct maple_node *node;
unsigned char offset;
unsigned long *pivots;
type = mte_node_type(mas->node);
node = mas_mn(mas);
if (type == maple_arange_64)
return ma_meta_end(node, type);
pivots = ma_pivots(node, type);
offset = mt_pivots[type] - 1;
if (likely(!pivots[offset]))
return ma_meta_end(node, type);
if (likely(pivots[offset] == mas->max))
return offset;
return mt_pivots[type];
}
/*
* mas_leaf_max_gap() - Returns the largest gap in a leaf node
* @mas - the maple state
*
* Return: The maximum gap in the leaf.
*/
static unsigned long mas_leaf_max_gap(struct ma_state *mas)
{
enum maple_type mt;
unsigned long pstart, gap, max_gap;
struct maple_node *mn;
unsigned long *pivots;
void __rcu **slots;
unsigned char i;
unsigned char max_piv;
mt = mte_node_type(mas->node);
mn = mas_mn(mas);
slots = ma_slots(mn, mt);
max_gap = 0;
if (unlikely(ma_is_dense(mt))) {
gap = 0;
for (i = 0; i < mt_slots[mt]; i++) {
if (slots[i]) {
if (gap > max_gap)
max_gap = gap;
gap = 0;
} else {
gap++;
}
}
if (gap > max_gap)
max_gap = gap;
return max_gap;
}
/*
* Check the first implied pivot optimizes the loop below and slot 1 may
* be skipped if there is a gap in slot 0.
*/
pivots = ma_pivots(mn, mt);
if (likely(!slots[0])) {
max_gap = pivots[0] - mas->min + 1;
i = 2;
} else {
i = 1;
}
/* reduce max_piv as the special case is checked before the loop */
max_piv = ma_data_end(mn, mt, pivots, mas->max) - 1;
/*
* Check end implied pivot which can only be a gap on the right most
* node.
*/
if (unlikely(mas->max == ULONG_MAX) && !slots[max_piv + 1]) {
gap = ULONG_MAX - pivots[max_piv];
if (gap > max_gap)
max_gap = gap;
}
for (; i <= max_piv; i++) {
/* data == no gap. */
if (likely(slots[i]))
continue;
pstart = pivots[i - 1];
gap = pivots[i] - pstart;
if (gap > max_gap)
max_gap = gap;
/* There cannot be two gaps in a row. */
i++;
}
return max_gap;
}
/*
* ma_max_gap() - Get the maximum gap in a maple node (non-leaf)
* @node: The maple node
* @gaps: The pointer to the gaps
* @mt: The maple node type
* @*off: Pointer to store the offset location of the gap.
*
* Uses the metadata data end to scan backwards across set gaps.
*
* Return: The maximum gap value
*/
static inline unsigned long
ma_max_gap(struct maple_node *node, unsigned long *gaps, enum maple_type mt,
unsigned char *off)
{
unsigned char offset, i;
unsigned long max_gap = 0;
i = offset = ma_meta_end(node, mt);
do {
if (gaps[i] > max_gap) {
max_gap = gaps[i];
offset = i;
}
} while (i--);
*off = offset;
return max_gap;
}
/*
* mas_max_gap() - find the largest gap in a non-leaf node and set the slot.
* @mas: The maple state.
*
* If the metadata gap is set to MAPLE_ARANGE64_META_MAX, there is no gap.
*
* Return: The gap value.
*/
static inline unsigned long mas_max_gap(struct ma_state *mas)
{
unsigned long *gaps;
unsigned char offset;
enum maple_type mt;
struct maple_node *node;
mt = mte_node_type(mas->node);
if (ma_is_leaf(mt))
return mas_leaf_max_gap(mas);
node = mas_mn(mas);
offset = ma_meta_gap(node, mt);
if (offset == MAPLE_ARANGE64_META_MAX)
return 0;
gaps = ma_gaps(node, mt);
return gaps[offset];
}
/*
* mas_parent_gap() - Set the parent gap and any gaps above, as needed
* @mas: The maple state
* @offset: The gap offset in the parent to set
* @new: The new gap value.
*
* Set the parent gap then continue to set the gap upwards, using the metadata
* of the parent to see if it is necessary to check the node above.
*/
static inline void mas_parent_gap(struct ma_state *mas, unsigned char offset,
unsigned long new)
{
unsigned long meta_gap = 0;
struct maple_node *pnode;
struct maple_enode *penode;
unsigned long *pgaps;
unsigned char meta_offset;
enum maple_type pmt;
pnode = mte_parent(mas->node);
pmt = mas_parent_enum(mas, mas->node);
penode = mt_mk_node(pnode, pmt);
pgaps = ma_gaps(pnode, pmt);
ascend:
meta_offset = ma_meta_gap(pnode, pmt);
if (meta_offset == MAPLE_ARANGE64_META_MAX)
meta_gap = 0;
else
meta_gap = pgaps[meta_offset];
pgaps[offset] = new;
if (meta_gap == new)
return;
if (offset != meta_offset) {
if (meta_gap > new)
return;
ma_set_meta_gap(pnode, pmt, offset);
} else if (new < meta_gap) {
meta_offset = 15;
new = ma_max_gap(pnode, pgaps, pmt, &meta_offset);
ma_set_meta_gap(pnode, pmt, meta_offset);
}
if (ma_is_root(pnode))
return;
/* Go to the parent node. */
pnode = mte_parent(penode);
pmt = mas_parent_enum(mas, penode);
pgaps = ma_gaps(pnode, pmt);
offset = mte_parent_slot(penode);
penode = mt_mk_node(pnode, pmt);
goto ascend;
}
/*
* mas_update_gap() - Update a nodes gaps and propagate up if necessary.
* @mas - the maple state.
*/
static inline void mas_update_gap(struct ma_state *mas)
{
unsigned char pslot;
unsigned long p_gap;
unsigned long max_gap;
if (!mt_is_alloc(mas->tree))
return;
if (mte_is_root(mas->node))
return;
max_gap = mas_max_gap(mas);
pslot = mte_parent_slot(mas->node);
p_gap = ma_gaps(mte_parent(mas->node),
mas_parent_enum(mas, mas->node))[pslot];
if (p_gap != max_gap)
mas_parent_gap(mas, pslot, max_gap);
}
/*
* mas_adopt_children() - Set the parent pointer of all nodes in @parent to
* @parent with the slot encoded.
* @mas - the maple state (for the tree)
* @parent - the maple encoded node containing the children.
*/
static inline void mas_adopt_children(struct ma_state *mas,
struct maple_enode *parent)
{
enum maple_type type = mte_node_type(parent);
struct maple_node *node = mas_mn(mas);
void __rcu **slots = ma_slots(node, type);
unsigned long *pivots = ma_pivots(node, type);
struct maple_enode *child;
unsigned char offset;
offset = ma_data_end(node, type, pivots, mas->max);
do {
child = mas_slot_locked(mas, slots, offset);
mte_set_parent(child, parent, offset);
} while (offset--);
}
/*
* mas_replace() - Replace a maple node in the tree with mas->node. Uses the
* parent encoding to locate the maple node in the tree.
* @mas - the ma_state to use for operations.
* @advanced - boolean to adopt the child nodes and free the old node (false) or
* leave the node (true) and handle the adoption and free elsewhere.
*/
static inline void mas_replace(struct ma_state *mas, bool advanced)
__must_hold(mas->tree->lock)
{
struct maple_node *mn = mas_mn(mas);
struct maple_enode *old_enode;
unsigned char offset = 0;
void __rcu **slots = NULL;
if (ma_is_root(mn)) {
old_enode = mas_root_locked(mas);
} else {
offset = mte_parent_slot(mas->node);
slots = ma_slots(mte_parent(mas->node),
mas_parent_enum(mas, mas->node));
old_enode = mas_slot_locked(mas, slots, offset);
}
if (!advanced && !mte_is_leaf(mas->node))
mas_adopt_children(mas, mas->node);
if (mte_is_root(mas->node)) {
mn->parent = ma_parent_ptr(
((unsigned long)mas->tree | MA_ROOT_PARENT));
rcu_assign_pointer(mas->tree->ma_root, mte_mk_root(mas->node));
mas_set_height(mas);
} else {
rcu_assign_pointer(slots[offset], mas->node);
}
if (!advanced)
mas_free(mas, old_enode);
}
/*
* mas_new_child() - Find the new child of a node.
* @mas: the maple state
* @child: the maple state to store the child.
*/
static inline bool mas_new_child(struct ma_state *mas, struct ma_state *child)
__must_hold(mas->tree->lock)
{
enum maple_type mt;
unsigned char offset;
unsigned char end;
unsigned long *pivots;
struct maple_enode *entry;
struct maple_node *node;
void __rcu **slots;
mt = mte_node_type(mas->node);
node = mas_mn(mas);
slots = ma_slots(node, mt);
pivots = ma_pivots(node, mt);
end = ma_data_end(node, mt, pivots, mas->max);
for (offset = mas->offset; offset <= end; offset++) {
entry = mas_slot_locked(mas, slots, offset);
if (mte_parent(entry) == node) {
*child = *mas;
mas->offset = offset + 1;
child->offset = offset;
mas_descend(child);
child->offset = 0;
return true;
}
}
return false;
}
/*
* mab_shift_right() - Shift the data in mab right. Note, does not clean out the
* old data or set b_node->b_end.
* @b_node: the maple_big_node
* @shift: the shift count
*/
static inline void mab_shift_right(struct maple_big_node *b_node,
unsigned char shift)
{
unsigned long size = b_node->b_end * sizeof(unsigned long);
memmove(b_node->pivot + shift, b_node->pivot, size);
memmove(b_node->slot + shift, b_node->slot, size);
if (b_node->type == maple_arange_64)
memmove(b_node->gap + shift, b_node->gap, size);
}
/*
* mab_middle_node() - Check if a middle node is needed (unlikely)
* @b_node: the maple_big_node that contains the data.
* @size: the amount of data in the b_node
* @split: the potential split location
* @slot_count: the size that can be stored in a single node being considered.
*
* Return: true if a middle node is required.
*/
static inline bool mab_middle_node(struct maple_big_node *b_node, int split,
unsigned char slot_count)
{
unsigned char size = b_node->b_end;
if (size >= 2 * slot_count)
return true;
if (!b_node->slot[split] && (size >= 2 * slot_count - 1))
return true;
return false;
}
/*
* mab_no_null_split() - ensure the split doesn't fall on a NULL
* @b_node: the maple_big_node with the data
* @split: the suggested split location
* @slot_count: the number of slots in the node being considered.
*
* Return: the split location.
*/
static inline int mab_no_null_split(struct maple_big_node *b_node,
unsigned char split, unsigned char slot_count)
{
if (!b_node->slot[split]) {
/*
* If the split is less than the max slot && the right side will
* still be sufficient, then increment the split on NULL.
*/
if ((split < slot_count - 1) &&
(b_node->b_end - split) > (mt_min_slots[b_node->type]))
split++;
else
split--;
}
return split;
}
/*
* mab_calc_split() - Calculate the split location and if there needs to be two
* splits.
* @bn: The maple_big_node with the data
* @mid_split: The second split, if required. 0 otherwise.
*
* Return: The first split location. The middle split is set in @mid_split.
*/
static inline int mab_calc_split(struct ma_state *mas,
struct maple_big_node *bn, unsigned char *mid_split, unsigned long min)
{
unsigned char b_end = bn->b_end;
int split = b_end / 2; /* Assume equal split. */
unsigned char slot_min, slot_count = mt_slots[bn->type];
/*
* To support gap tracking, all NULL entries are kept together and a node cannot
* end on a NULL entry, with the exception of the left-most leaf. The
* limitation means that the split of a node must be checked for this condition
* and be able to put more data in one direction or the other.
*/
if (unlikely((mas->mas_flags & MA_STATE_BULK))) {
*mid_split = 0;
split = b_end - mt_min_slots[bn->type];
if (!ma_is_leaf(bn->type))
return split;
mas->mas_flags |= MA_STATE_REBALANCE;
if (!bn->slot[split])
split--;
return split;
}
/*
* Although extremely rare, it is possible to enter what is known as the 3-way
* split scenario. The 3-way split comes about by means of a store of a range
* that overwrites the end and beginning of two full nodes. The result is a set
* of entries that cannot be stored in 2 nodes. Sometimes, these two nodes can
* also be located in different parent nodes which are also full. This can
* carry upwards all the way to the root in the worst case.
*/
if (unlikely(mab_middle_node(bn, split, slot_count))) {
split = b_end / 3;
*mid_split = split * 2;
} else {
slot_min = mt_min_slots[bn->type];
*mid_split = 0;
/*
* Avoid having a range less than the slot count unless it
* causes one node to be deficient.
* NOTE: mt_min_slots is 1 based, b_end and split are zero.
*/
while (((bn->pivot[split] - min) < slot_count - 1) &&
(split < slot_count - 1) && (b_end - split > slot_min))
split++;
}
/* Avoid ending a node on a NULL entry */
split = mab_no_null_split(bn, split, slot_count);
if (unlikely(*mid_split))
*mid_split = mab_no_null_split(bn, *mid_split, slot_count);
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
return split;
}
/*
* mas_mab_cp() - Copy data from a maple state inclusively to a maple_big_node
* and set @b_node->b_end to the next free slot.
* @mas: The maple state
* @mas_start: The starting slot to copy
* @mas_end: The end slot to copy (inclusively)
* @b_node: The maple_big_node to place the data
* @mab_start: The starting location in maple_big_node to store the data.
*/
static inline void mas_mab_cp(struct ma_state *mas, unsigned char mas_start,
unsigned char mas_end, struct maple_big_node *b_node,
unsigned char mab_start)
{
enum maple_type mt;
struct maple_node *node;
void __rcu **slots;
unsigned long *pivots, *gaps;
int i = mas_start, j = mab_start;
unsigned char piv_end;
node = mas_mn(mas);
mt = mte_node_type(mas->node);
pivots = ma_pivots(node, mt);
if (!i) {
b_node->pivot[j] = pivots[i++];
if (unlikely(i > mas_end))
goto complete;
j++;
}
piv_end = min(mas_end, mt_pivots[mt]);
for (; i < piv_end; i++, j++) {
b_node->pivot[j] = pivots[i];
if (unlikely(!b_node->pivot[j]))
break;
if (unlikely(mas->max == b_node->pivot[j]))
goto complete;
}
if (likely(i <= mas_end))
b_node->pivot[j] = mas_safe_pivot(mas, pivots, i, mt);
complete:
b_node->b_end = ++j;
j -= mab_start;
slots = ma_slots(node, mt);
memcpy(b_node->slot + mab_start, slots + mas_start, sizeof(void *) * j);
if (!ma_is_leaf(mt) && mt_is_alloc(mas->tree)) {
gaps = ma_gaps(node, mt);
memcpy(b_node->gap + mab_start, gaps + mas_start,
sizeof(unsigned long) * j);
}
}
/*
* mas_leaf_set_meta() - Set the metadata of a leaf if possible.
* @mas: The maple state
* @node: The maple node
* @pivots: pointer to the maple node pivots
* @mt: The maple type
* @end: The assumed end
*
* Note, end may be incremented within this function but not modified at the
* source. This is fine since the metadata is the last thing to be stored in a
* node during a write.
*/
static inline void mas_leaf_set_meta(struct ma_state *mas,
struct maple_node *node, unsigned long *pivots,
enum maple_type mt, unsigned char end)
{
/* There is no room for metadata already */
if (mt_pivots[mt] <= end)
return;
if (pivots[end] && pivots[end] < mas->max)
end++;
if (end < mt_slots[mt] - 1)
ma_set_meta(node, mt, 0, end);
}
/*
* mab_mas_cp() - Copy data from maple_big_node to a maple encoded node.
* @b_node: the maple_big_node that has the data
* @mab_start: the start location in @b_node.
* @mab_end: The end location in @b_node (inclusively)
* @mas: The maple state with the maple encoded node.
*/
static inline void mab_mas_cp(struct maple_big_node *b_node,
unsigned char mab_start, unsigned char mab_end,
struct ma_state *mas, bool new_max)
{
int i, j = 0;
enum maple_type mt = mte_node_type(mas->node);
struct maple_node *node = mte_to_node(mas->node);
void __rcu **slots = ma_slots(node, mt);
unsigned long *pivots = ma_pivots(node, mt);
unsigned long *gaps = NULL;
unsigned char end;
if (mab_end - mab_start > mt_pivots[mt])
mab_end--;
if (!pivots[mt_pivots[mt] - 1])
slots[mt_pivots[mt]] = NULL;
i = mab_start;
do {
pivots[j++] = b_node->pivot[i++];
} while (i <= mab_end && likely(b_node->pivot[i]));
memcpy(slots, b_node->slot + mab_start,
sizeof(void *) * (i - mab_start));
if (new_max)
mas->max = b_node->pivot[i - 1];
end = j - 1;
if (likely(!ma_is_leaf(mt) && mt_is_alloc(mas->tree))) {
unsigned long max_gap = 0;
unsigned char offset = 15;
gaps = ma_gaps(node, mt);
do {
gaps[--j] = b_node->gap[--i];
if (gaps[j] > max_gap) {
offset = j;
max_gap = gaps[j];
}
} while (j);
ma_set_meta(node, mt, offset, end);
} else {
mas_leaf_set_meta(mas, node, pivots, mt, end);
}
}
/*
* mas_descend_adopt() - Descend through a sub-tree and adopt children.
* @mas: the maple state with the maple encoded node of the sub-tree.
*
* Descend through a sub-tree and adopt children who do not have the correct
* parents set. Follow the parents which have the correct parents as they are
* the new entries which need to be followed to find other incorrectly set
* parents.
*/
static inline void mas_descend_adopt(struct ma_state *mas)
{
struct ma_state list[3], next[3];
int i, n;
/*
* At each level there may be up to 3 correct parent pointers which indicates
* the new nodes which need to be walked to find any new nodes at a lower level.
*/
for (i = 0; i < 3; i++) {
list[i] = *mas;
list[i].offset = 0;
next[i].offset = 0;
}
next[0] = *mas;
while (!mte_is_leaf(list[0].node)) {
n = 0;
for (i = 0; i < 3; i++) {
if (mas_is_none(&list[i]))
continue;
if (i && list[i-1].node == list[i].node)
continue;
while ((n < 3) && (mas_new_child(&list[i], &next[n])))
n++;
mas_adopt_children(&list[i], list[i].node);
}
while (n < 3)
next[n++].node = MAS_NONE;
/* descend by setting the list to the children */
for (i = 0; i < 3; i++)
list[i] = next[i];
}
}
/*
* mas_bulk_rebalance() - Rebalance the end of a tree after a bulk insert.
* @mas: The maple state
* @end: The maple node end
* @mt: The maple node type
*/
static inline void mas_bulk_rebalance(struct ma_state *mas, unsigned char end,
enum maple_type mt)
{
if (!(mas->mas_flags & MA_STATE_BULK))
return;
if (mte_is_root(mas->node))
return;
if (end > mt_min_slots[mt]) {
mas->mas_flags &= ~MA_STATE_REBALANCE;
return;
}
}
/*
* mas_store_b_node() - Store an @entry into the b_node while also copying the
* data from a maple encoded node.
* @wr_mas: the maple write state
* @b_node: the maple_big_node to fill with data
* @offset_end: the offset to end copying
*
* Return: The actual end of the data stored in @b_node
*/
static noinline_for_kasan void mas_store_b_node(struct ma_wr_state *wr_mas,
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
struct maple_big_node *b_node, unsigned char offset_end)
{
unsigned char slot;
unsigned char b_end;
/* Possible underflow of piv will wrap back to 0 before use. */
unsigned long piv;
struct ma_state *mas = wr_mas->mas;
b_node->type = wr_mas->type;
b_end = 0;
slot = mas->offset;
if (slot) {
/* Copy start data up to insert. */
mas_mab_cp(mas, 0, slot - 1, b_node, 0);
b_end = b_node->b_end;
piv = b_node->pivot[b_end - 1];
} else
piv = mas->min - 1;
if (piv + 1 < mas->index) {
/* Handle range starting after old range */
b_node->slot[b_end] = wr_mas->content;
if (!wr_mas->content)
b_node->gap[b_end] = mas->index - 1 - piv;
b_node->pivot[b_end++] = mas->index - 1;
}
/* Store the new entry. */
mas->offset = b_end;
b_node->slot[b_end] = wr_mas->entry;
b_node->pivot[b_end] = mas->last;
/* Appended. */
if (mas->last >= mas->max)
goto b_end;
/* Handle new range ending before old range ends */
piv = mas_logical_pivot(mas, wr_mas->pivots, offset_end, wr_mas->type);
if (piv > mas->last) {
if (piv == ULONG_MAX)
mas_bulk_rebalance(mas, b_node->b_end, wr_mas->type);
if (offset_end != slot)
wr_mas->content = mas_slot_locked(mas, wr_mas->slots,
offset_end);
b_node->slot[++b_end] = wr_mas->content;
if (!wr_mas->content)
b_node->gap[b_end] = piv - mas->last + 1;
b_node->pivot[b_end] = piv;
}
slot = offset_end + 1;
if (slot > wr_mas->node_end)
goto b_end;
/* Copy end data to the end of the node. */
mas_mab_cp(mas, slot, wr_mas->node_end + 1, b_node, ++b_end);
b_node->b_end--;
return;
b_end:
b_node->b_end = b_end;
}
/*
* mas_prev_sibling() - Find the previous node with the same parent.
* @mas: the maple state
*
* Return: True if there is a previous sibling, false otherwise.
*/
static inline bool mas_prev_sibling(struct ma_state *mas)
{
unsigned int p_slot = mte_parent_slot(mas->node);
if (mte_is_root(mas->node))
return false;
if (!p_slot)
return false;
mas_ascend(mas);
mas->offset = p_slot - 1;
mas_descend(mas);
return true;
}
/*
* mas_next_sibling() - Find the next node with the same parent.
* @mas: the maple state
*
* Return: true if there is a next sibling, false otherwise.
*/
static inline bool mas_next_sibling(struct ma_state *mas)
{
MA_STATE(parent, mas->tree, mas->index, mas->last);
if (mte_is_root(mas->node))
return false;
parent = *mas;
mas_ascend(&parent);
parent.offset = mte_parent_slot(mas->node) + 1;
if (parent.offset > mas_data_end(&parent))
return false;
*mas = parent;
mas_descend(mas);
return true;
}
/*
* mte_node_or_node() - Return the encoded node or MAS_NONE.
* @enode: The encoded maple node.
*
* Shorthand to avoid setting %NULLs in the tree or maple_subtree_state.
*
* Return: @enode or MAS_NONE
*/
static inline struct maple_enode *mte_node_or_none(struct maple_enode *enode)
{
if (enode)
return enode;
return ma_enode_ptr(MAS_NONE);
}
/*
* mas_wr_node_walk() - Find the correct offset for the index in the @mas.
* @wr_mas: The maple write state
*
* Uses mas_slot_locked() and does not need to worry about dead nodes.
*/
static inline void mas_wr_node_walk(struct ma_wr_state *wr_mas)
{
struct ma_state *mas = wr_mas->mas;
unsigned char count;
unsigned char offset;
unsigned long index, min, max;
if (unlikely(ma_is_dense(wr_mas->type))) {
wr_mas->r_max = wr_mas->r_min = mas->index;
mas->offset = mas->index = mas->min;
return;
}
wr_mas->node = mas_mn(wr_mas->mas);
wr_mas->pivots = ma_pivots(wr_mas->node, wr_mas->type);
count = wr_mas->node_end = ma_data_end(wr_mas->node, wr_mas->type,
wr_mas->pivots, mas->max);
offset = mas->offset;
min = mas_safe_min(mas, wr_mas->pivots, offset);
if (unlikely(offset == count))
goto max;
max = wr_mas->pivots[offset];
index = mas->index;
if (unlikely(index <= max))
goto done;
if (unlikely(!max && offset))
goto max;
min = max + 1;
while (++offset < count) {
max = wr_mas->pivots[offset];
if (index <= max)
goto done;
else if (unlikely(!max))
break;
min = max + 1;
}
max:
max = mas->max;
done:
wr_mas->r_max = max;
wr_mas->r_min = min;
wr_mas->offset_end = mas->offset = offset;
}
/*
* mas_topiary_range() - Add a range of slots to the topiary.
* @mas: The maple state
* @destroy: The topiary to add the slots (usually destroy)
* @start: The starting slot inclusively
* @end: The end slot inclusively
*/
static inline void mas_topiary_range(struct ma_state *mas,
struct ma_topiary *destroy, unsigned char start, unsigned char end)
{
void __rcu **slots;
unsigned char offset;
MT_BUG_ON(mas->tree, mte_is_leaf(mas->node));
slots = ma_slots(mas_mn(mas), mte_node_type(mas->node));
for (offset = start; offset <= end; offset++) {
struct maple_enode *enode = mas_slot_locked(mas, slots, offset);
if (mte_dead_node(enode))
continue;
mat_add(destroy, enode);
}
}
/*
* mast_topiary() - Add the portions of the tree to the removal list; either to
* be freed or discarded (destroy walk).
* @mast: The maple_subtree_state.
*/
static inline void mast_topiary(struct maple_subtree_state *mast)
{
MA_WR_STATE(wr_mas, mast->orig_l, NULL);
unsigned char r_start, r_end;
unsigned char l_start, l_end;
void __rcu **l_slots, **r_slots;
wr_mas.type = mte_node_type(mast->orig_l->node);
mast->orig_l->index = mast->orig_l->last;
mas_wr_node_walk(&wr_mas);
l_start = mast->orig_l->offset + 1;
l_end = mas_data_end(mast->orig_l);
r_start = 0;
r_end = mast->orig_r->offset;
if (r_end)
r_end--;
l_slots = ma_slots(mas_mn(mast->orig_l),
mte_node_type(mast->orig_l->node));
r_slots = ma_slots(mas_mn(mast->orig_r),
mte_node_type(mast->orig_r->node));
if ((l_start < l_end) &&
mte_dead_node(mas_slot_locked(mast->orig_l, l_slots, l_start))) {
l_start++;
}
if (mte_dead_node(mas_slot_locked(mast->orig_r, r_slots, r_end))) {
if (r_end)
r_end--;
}
if ((l_start > r_end) && (mast->orig_l->node == mast->orig_r->node))
return;
/* At the node where left and right sides meet, add the parts between */
if (mast->orig_l->node == mast->orig_r->node) {
return mas_topiary_range(mast->orig_l, mast->destroy,
l_start, r_end);
}
/* mast->orig_r is different and consumed. */
if (mte_is_leaf(mast->orig_r->node))
return;
if (mte_dead_node(mas_slot_locked(mast->orig_l, l_slots, l_end)))
l_end--;
if (l_start <= l_end)
mas_topiary_range(mast->orig_l, mast->destroy, l_start, l_end);
if (mte_dead_node(mas_slot_locked(mast->orig_r, r_slots, r_start)))
r_start++;
if (r_start <= r_end)
mas_topiary_range(mast->orig_r, mast->destroy, 0, r_end);
}
/*
* mast_rebalance_next() - Rebalance against the next node
* @mast: The maple subtree state
* @old_r: The encoded maple node to the right (next node).
*/
static inline void mast_rebalance_next(struct maple_subtree_state *mast)
{
unsigned char b_end = mast->bn->b_end;
mas_mab_cp(mast->orig_r, 0, mt_slot_count(mast->orig_r->node),
mast->bn, b_end);
mast->orig_r->last = mast->orig_r->max;
}
/*
* mast_rebalance_prev() - Rebalance against the previous node
* @mast: The maple subtree state
* @old_l: The encoded maple node to the left (previous node)
*/
static inline void mast_rebalance_prev(struct maple_subtree_state *mast)
{
unsigned char end = mas_data_end(mast->orig_l) + 1;
unsigned char b_end = mast->bn->b_end;
mab_shift_right(mast->bn, end);
mas_mab_cp(mast->orig_l, 0, end - 1, mast->bn, 0);
mast->l->min = mast->orig_l->min;
mast->orig_l->index = mast->orig_l->min;
mast->bn->b_end = end + b_end;
mast->l->offset += end;
}
/*
* mast_spanning_rebalance() - Rebalance nodes with nearest neighbour favouring
* the node to the right. Checking the nodes to the right then the left at each
* level upwards until root is reached. Free and destroy as needed.
* Data is copied into the @mast->bn.
* @mast: The maple_subtree_state.
*/
static inline
bool mast_spanning_rebalance(struct maple_subtree_state *mast)
{
struct ma_state r_tmp = *mast->orig_r;
struct ma_state l_tmp = *mast->orig_l;
struct maple_enode *ancestor = NULL;
unsigned char start, end;
unsigned char depth = 0;
r_tmp = *mast->orig_r;
l_tmp = *mast->orig_l;
do {
mas_ascend(mast->orig_r);
mas_ascend(mast->orig_l);
depth++;
if (!ancestor &&
(mast->orig_r->node == mast->orig_l->node)) {
ancestor = mast->orig_r->node;
end = mast->orig_r->offset - 1;
start = mast->orig_l->offset + 1;
}
if (mast->orig_r->offset < mas_data_end(mast->orig_r)) {
if (!ancestor) {
ancestor = mast->orig_r->node;
start = 0;
}
mast->orig_r->offset++;
do {
mas_descend(mast->orig_r);
mast->orig_r->offset = 0;
depth--;
} while (depth);
mast_rebalance_next(mast);
do {
unsigned char l_off = 0;
struct maple_enode *child = r_tmp.node;
mas_ascend(&r_tmp);
if (ancestor == r_tmp.node)
l_off = start;
if (r_tmp.offset)
r_tmp.offset--;
if (l_off < r_tmp.offset)
mas_topiary_range(&r_tmp, mast->destroy,
l_off, r_tmp.offset);
if (l_tmp.node != child)
mat_add(mast->free, child);
} while (r_tmp.node != ancestor);
*mast->orig_l = l_tmp;
return true;
} else if (mast->orig_l->offset != 0) {
if (!ancestor) {
ancestor = mast->orig_l->node;
end = mas_data_end(mast->orig_l);
}
mast->orig_l->offset--;
do {
mas_descend(mast->orig_l);
mast->orig_l->offset =
mas_data_end(mast->orig_l);
depth--;
} while (depth);
mast_rebalance_prev(mast);
do {
unsigned char r_off;
struct maple_enode *child = l_tmp.node;
mas_ascend(&l_tmp);
if (ancestor == l_tmp.node)
r_off = end;
else
r_off = mas_data_end(&l_tmp);
if (l_tmp.offset < r_off)
l_tmp.offset++;
if (l_tmp.offset < r_off)
mas_topiary_range(&l_tmp, mast->destroy,
l_tmp.offset, r_off);
if (r_tmp.node != child)
mat_add(mast->free, child);
} while (l_tmp.node != ancestor);
*mast->orig_r = r_tmp;
return true;
}
} while (!mte_is_root(mast->orig_r->node));
*mast->orig_r = r_tmp;
*mast->orig_l = l_tmp;
return false;
}
/*
* mast_ascend_free() - Add current original maple state nodes to the free list
* and ascend.
* @mast: the maple subtree state.
*
* Ascend the original left and right sides and add the previous nodes to the
* free list. Set the slots to point to the correct location in the new nodes.
*/
static inline void
mast_ascend_free(struct maple_subtree_state *mast)
{
MA_WR_STATE(wr_mas, mast->orig_r, NULL);
struct maple_enode *left = mast->orig_l->node;
struct maple_enode *right = mast->orig_r->node;
mas_ascend(mast->orig_l);
mas_ascend(mast->orig_r);
mat_add(mast->free, left);
if (left != right)
mat_add(mast->free, right);
mast->orig_r->offset = 0;
mast->orig_r->index = mast->r->max;
/* last should be larger than or equal to index */
if (mast->orig_r->last < mast->orig_r->index)
mast->orig_r->last = mast->orig_r->index;
/*
* The node may not contain the value so set slot to ensure all
* of the nodes contents are freed or destroyed.
*/
wr_mas.type = mte_node_type(mast->orig_r->node);
mas_wr_node_walk(&wr_mas);
/* Set up the left side of things */
mast->orig_l->offset = 0;
mast->orig_l->index = mast->l->min;
wr_mas.mas = mast->orig_l;
wr_mas.type = mte_node_type(mast->orig_l->node);
mas_wr_node_walk(&wr_mas);
mast->bn->type = wr_mas.type;
}
/*
* mas_new_ma_node() - Create and return a new maple node. Helper function.
* @mas: the maple state with the allocations.
* @b_node: the maple_big_node with the type encoding.
*
* Use the node type from the maple_big_node to allocate a new node from the
* ma_state. This function exists mainly for code readability.
*
* Return: A new maple encoded node
*/
static inline struct maple_enode
*mas_new_ma_node(struct ma_state *mas, struct maple_big_node *b_node)
{
return mt_mk_node(ma_mnode_ptr(mas_pop_node(mas)), b_node->type);
}
/*
* mas_mab_to_node() - Set up right and middle nodes
*
* @mas: the maple state that contains the allocations.
* @b_node: the node which contains the data.
* @left: The pointer which will have the left node
* @right: The pointer which may have the right node
* @middle: the pointer which may have the middle node (rare)
* @mid_split: the split location for the middle node
*
* Return: the split of left.
*/
static inline unsigned char mas_mab_to_node(struct ma_state *mas,
struct maple_big_node *b_node, struct maple_enode **left,
struct maple_enode **right, struct maple_enode **middle,
unsigned char *mid_split, unsigned long min)
{
unsigned char split = 0;
unsigned char slot_count = mt_slots[b_node->type];
*left = mas_new_ma_node(mas, b_node);
*right = NULL;
*middle = NULL;
*mid_split = 0;
if (b_node->b_end < slot_count) {
split = b_node->b_end;
} else {
split = mab_calc_split(mas, b_node, mid_split, min);
*right = mas_new_ma_node(mas, b_node);
}
if (*mid_split)
*middle = mas_new_ma_node(mas, b_node);
return split;
}
/*
* mab_set_b_end() - Add entry to b_node at b_node->b_end and increment the end
* pointer.
* @b_node - the big node to add the entry
* @mas - the maple state to get the pivot (mas->max)
* @entry - the entry to add, if NULL nothing happens.
*/
static inline void mab_set_b_end(struct maple_big_node *b_node,
struct ma_state *mas,
void *entry)
{
if (!entry)
return;
b_node->slot[b_node->b_end] = entry;
if (mt_is_alloc(mas->tree))
b_node->gap[b_node->b_end] = mas_max_gap(mas);
b_node->pivot[b_node->b_end++] = mas->max;
}
/*
* mas_set_split_parent() - combine_then_separate helper function. Sets the parent
* of @mas->node to either @left or @right, depending on @slot and @split
*
* @mas - the maple state with the node that needs a parent
* @left - possible parent 1
* @right - possible parent 2
* @slot - the slot the mas->node was placed
* @split - the split location between @left and @right
*/
static inline void mas_set_split_parent(struct ma_state *mas,
struct maple_enode *left,
struct maple_enode *right,
unsigned char *slot, unsigned char split)
{
if (mas_is_none(mas))
return;
if ((*slot) <= split)
mte_set_parent(mas->node, left, *slot);
else if (right)
mte_set_parent(mas->node, right, (*slot) - split - 1);
(*slot)++;
}
/*
* mte_mid_split_check() - Check if the next node passes the mid-split
* @**l: Pointer to left encoded maple node.
* @**m: Pointer to middle encoded maple node.
* @**r: Pointer to right encoded maple node.
* @slot: The offset
* @*split: The split location.
* @mid_split: The middle split.
*/
static inline void mte_mid_split_check(struct maple_enode **l,
struct maple_enode **r,
struct maple_enode *right,
unsigned char slot,
unsigned char *split,
unsigned char mid_split)
{
if (*r == right)
return;
if (slot < mid_split)
return;
*l = *r;
*r = right;
*split = mid_split;
}
/*
* mast_set_split_parents() - Helper function to set three nodes parents. Slot
* is taken from @mast->l.
* @mast - the maple subtree state
* @left - the left node
* @right - the right node
* @split - the split location.
*/
static inline void mast_set_split_parents(struct maple_subtree_state *mast,
struct maple_enode *left,
struct maple_enode *middle,
struct maple_enode *right,
unsigned char split,
unsigned char mid_split)
{
unsigned char slot;
struct maple_enode *l = left;
struct maple_enode *r = right;
if (mas_is_none(mast->l))
return;
if (middle)
r = middle;
slot = mast->l->offset;
mte_mid_split_check(&l, &r, right, slot, &split, mid_split);
mas_set_split_parent(mast->l, l, r, &slot, split);
mte_mid_split_check(&l, &r, right, slot, &split, mid_split);
mas_set_split_parent(mast->m, l, r, &slot, split);
mte_mid_split_check(&l, &r, right, slot, &split, mid_split);
mas_set_split_parent(mast->r, l, r, &slot, split);
}
/*
* mas_wmb_replace() - Write memory barrier and replace
* @mas: The maple state
* @free: the maple topiary list of nodes to free
* @destroy: The maple topiary list of nodes to destroy (walk and free)
*
* Updates gap as necessary.
*/
static inline void mas_wmb_replace(struct ma_state *mas,
struct ma_topiary *free,
struct ma_topiary *destroy)
{
/* All nodes must see old data as dead prior to replacing that data */
smp_wmb(); /* Needed for RCU */
/* Insert the new data in the tree */
mas_replace(mas, true);
if (!mte_is_leaf(mas->node))
mas_descend_adopt(mas);
mas_mat_free(mas, free);
if (destroy)
mas_mat_destroy(mas, destroy);
if (mte_is_leaf(mas->node))
return;
mas_update_gap(mas);
}
/*
* mast_new_root() - Set a new tree root during subtree creation
* @mast: The maple subtree state
* @mas: The maple state
*/
static inline void mast_new_root(struct maple_subtree_state *mast,
struct ma_state *mas)
{
mas_mn(mast->l)->parent =
ma_parent_ptr(((unsigned long)mas->tree | MA_ROOT_PARENT));
if (!mte_dead_node(mast->orig_l->node) &&
!mte_is_root(mast->orig_l->node)) {
do {
mast_ascend_free(mast);
mast_topiary(mast);
} while (!mte_is_root(mast->orig_l->node));
}
if ((mast->orig_l->node != mas->node) &&
(mast->l->depth > mas_mt_height(mas))) {
mat_add(mast->free, mas->node);
}
}
/*
* mast_cp_to_nodes() - Copy data out to nodes.
* @mast: The maple subtree state
* @left: The left encoded maple node
* @middle: The middle encoded maple node
* @right: The right encoded maple node
* @split: The location to split between left and (middle ? middle : right)
* @mid_split: The location to split between middle and right.
*/
static inline void mast_cp_to_nodes(struct maple_subtree_state *mast,
struct maple_enode *left, struct maple_enode *middle,
struct maple_enode *right, unsigned char split, unsigned char mid_split)
{
bool new_lmax = true;
mast->l->node = mte_node_or_none(left);
mast->m->node = mte_node_or_none(middle);
mast->r->node = mte_node_or_none(right);
mast->l->min = mast->orig_l->min;
if (split == mast->bn->b_end) {
mast->l->max = mast->orig_r->max;
new_lmax = false;
}
mab_mas_cp(mast->bn, 0, split, mast->l, new_lmax);
if (middle) {
mab_mas_cp(mast->bn, 1 + split, mid_split, mast->m, true);
mast->m->min = mast->bn->pivot[split] + 1;
split = mid_split;
}
mast->r->max = mast->orig_r->max;
if (right) {
mab_mas_cp(mast->bn, 1 + split, mast->bn->b_end, mast->r, false);
mast->r->min = mast->bn->pivot[split] + 1;
}
}
/*
* mast_combine_cp_left - Copy in the original left side of the tree into the
* combined data set in the maple subtree state big node.
* @mast: The maple subtree state
*/
static inline void mast_combine_cp_left(struct maple_subtree_state *mast)
{
unsigned char l_slot = mast->orig_l->offset;
if (!l_slot)
return;
mas_mab_cp(mast->orig_l, 0, l_slot - 1, mast->bn, 0);
}
/*
* mast_combine_cp_right: Copy in the original right side of the tree into the
* combined data set in the maple subtree state big node.
* @mast: The maple subtree state
*/
static inline void mast_combine_cp_right(struct maple_subtree_state *mast)
{
if (mast->bn->pivot[mast->bn->b_end - 1] >= mast->orig_r->max)
return;
mas_mab_cp(mast->orig_r, mast->orig_r->offset + 1,
mt_slot_count(mast->orig_r->node), mast->bn,
mast->bn->b_end);
mast->orig_r->last = mast->orig_r->max;
}
/*
* mast_sufficient: Check if the maple subtree state has enough data in the big
* node to create at least one sufficient node
* @mast: the maple subtree state
*/
static inline bool mast_sufficient(struct maple_subtree_state *mast)
{
if (mast->bn->b_end > mt_min_slot_count(mast->orig_l->node))
return true;
return false;
}
/*
* mast_overflow: Check if there is too much data in the subtree state for a
* single node.
* @mast: The maple subtree state
*/
static inline bool mast_overflow(struct maple_subtree_state *mast)
{
if (mast->bn->b_end >= mt_slot_count(mast->orig_l->node))
return true;
return false;
}
static inline void *mtree_range_walk(struct ma_state *mas)
{
unsigned long *pivots;
unsigned char offset;
struct maple_node *node;
struct maple_enode *next, *last;
enum maple_type type;
void __rcu **slots;
unsigned char end;
unsigned long max, min;
unsigned long prev_max, prev_min;
next = mas->node;
min = mas->min;
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
max = mas->max;
do {
offset = 0;
last = next;
node = mte_to_node(next);
type = mte_node_type(next);
pivots = ma_pivots(node, type);
end = ma_data_end(node, type, pivots, max);
if (unlikely(ma_dead_node(node)))
goto dead_node;
if (pivots[offset] >= mas->index) {
prev_max = max;
prev_min = min;
max = pivots[offset];
goto next;
}
do {
offset++;
} while ((offset < end) && (pivots[offset] < mas->index));
prev_min = min;
min = pivots[offset - 1] + 1;
prev_max = max;
if (likely(offset < end && pivots[offset]))
max = pivots[offset];
next:
slots = ma_slots(node, type);
next = mt_slot(mas->tree, slots, offset);
if (unlikely(ma_dead_node(node)))
goto dead_node;
} while (!ma_is_leaf(type));
mas->offset = offset;
mas->index = min;
mas->last = max;
mas->min = prev_min;
mas->max = prev_max;
mas->node = last;
return (void *)next;
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
dead_node:
mas_reset(mas);
return NULL;
}
/*
* mas_spanning_rebalance() - Rebalance across two nodes which may not be peers.
* @mas: The starting maple state
* @mast: The maple_subtree_state, keeps track of 4 maple states.
* @count: The estimated count of iterations needed.
*
* Follow the tree upwards from @l_mas and @r_mas for @count, or until the root
* is hit. First @b_node is split into two entries which are inserted into the
* next iteration of the loop. @b_node is returned populated with the final
* iteration. @mas is used to obtain allocations. orig_l_mas keeps track of the
* nodes that will remain active by using orig_l_mas->index and orig_l_mas->last
* to account of what has been copied into the new sub-tree. The update of
* orig_l_mas->last is used in mas_consume to find the slots that will need to
* be either freed or destroyed. orig_l_mas->depth keeps track of the height of
* the new sub-tree in case the sub-tree becomes the full tree.
*
* Return: the number of elements in b_node during the last loop.
*/
static int mas_spanning_rebalance(struct ma_state *mas,
struct maple_subtree_state *mast, unsigned char count)
{
unsigned char split, mid_split;
unsigned char slot = 0;
struct maple_enode *left = NULL, *middle = NULL, *right = NULL;
MA_STATE(l_mas, mas->tree, mas->index, mas->index);
MA_STATE(r_mas, mas->tree, mas->index, mas->last);
MA_STATE(m_mas, mas->tree, mas->index, mas->index);
MA_TOPIARY(free, mas->tree);
MA_TOPIARY(destroy, mas->tree);
/*
* The tree needs to be rebalanced and leaves need to be kept at the same level.
* Rebalancing is done by use of the ``struct maple_topiary``.
*/
mast->l = &l_mas;
mast->m = &m_mas;
mast->r = &r_mas;
mast->free = &free;
mast->destroy = &destroy;
l_mas.node = r_mas.node = m_mas.node = MAS_NONE;
/* Check if this is not root and has sufficient data. */
if (((mast->orig_l->min != 0) || (mast->orig_r->max != ULONG_MAX)) &&
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
unlikely(mast->bn->b_end <= mt_min_slots[mast->bn->type]))
mast_spanning_rebalance(mast);
mast->orig_l->depth = 0;
/*
* Each level of the tree is examined and balanced, pushing data to the left or
* right, or rebalancing against left or right nodes is employed to avoid
* rippling up the tree to limit the amount of churn. Once a new sub-section of
* the tree is created, there may be a mix of new and old nodes. The old nodes
* will have the incorrect parent pointers and currently be in two trees: the
* original tree and the partially new tree. To remedy the parent pointers in
* the old tree, the new data is swapped into the active tree and a walk down
* the tree is performed and the parent pointers are updated.
* See mas_descend_adopt() for more information..
*/
while (count--) {
mast->bn->b_end--;
mast->bn->type = mte_node_type(mast->orig_l->node);
split = mas_mab_to_node(mas, mast->bn, &left, &right, &middle,
&mid_split, mast->orig_l->min);
mast_set_split_parents(mast, left, middle, right, split,
mid_split);
mast_cp_to_nodes(mast, left, middle, right, split, mid_split);
/*
* Copy data from next level in the tree to mast->bn from next
* iteration
*/
memset(mast->bn, 0, sizeof(struct maple_big_node));
mast->bn->type = mte_node_type(left);
mast->orig_l->depth++;
/* Root already stored in l->node. */
if (mas_is_root_limits(mast->l))
goto new_root;
mast_ascend_free(mast);
mast_combine_cp_left(mast);
l_mas.offset = mast->bn->b_end;
mab_set_b_end(mast->bn, &l_mas, left);
mab_set_b_end(mast->bn, &m_mas, middle);
mab_set_b_end(mast->bn, &r_mas, right);
/* Copy anything necessary out of the right node. */
mast_combine_cp_right(mast);
mast_topiary(mast);
mast->orig_l->last = mast->orig_l->max;
if (mast_sufficient(mast))
continue;
if (mast_overflow(mast))
continue;
/* May be a new root stored in mast->bn */
if (mas_is_root_limits(mast->orig_l))
break;
mast_spanning_rebalance(mast);
/* rebalancing from other nodes may require another loop. */
if (!count)
count++;
}
l_mas.node = mt_mk_node(ma_mnode_ptr(mas_pop_node(mas)),
mte_node_type(mast->orig_l->node));
mast->orig_l->depth++;
mab_mas_cp(mast->bn, 0, mt_slots[mast->bn->type] - 1, &l_mas, true);
mte_set_parent(left, l_mas.node, slot);
if (middle)
mte_set_parent(middle, l_mas.node, ++slot);
if (right)
mte_set_parent(right, l_mas.node, ++slot);
if (mas_is_root_limits(mast->l)) {
new_root:
mast_new_root(mast, mas);
} else {
mas_mn(&l_mas)->parent = mas_mn(mast->orig_l)->parent;
}
if (!mte_dead_node(mast->orig_l->node))
mat_add(&free, mast->orig_l->node);
mas->depth = mast->orig_l->depth;
*mast->orig_l = l_mas;
mte_set_node_dead(mas->node);
/* Set up mas for insertion. */
mast->orig_l->depth = mas->depth;
mast->orig_l->alloc = mas->alloc;
*mas = *mast->orig_l;
mas_wmb_replace(mas, &free, &destroy);
mtree_range_walk(mas);
return mast->bn->b_end;
}
/*
* mas_rebalance() - Rebalance a given node.
* @mas: The maple state
* @b_node: The big maple node.
*
* Rebalance two nodes into a single node or two new nodes that are sufficient.
* Continue upwards until tree is sufficient.
*
* Return: the number of elements in b_node during the last loop.
*/
static inline int mas_rebalance(struct ma_state *mas,
struct maple_big_node *b_node)
{
char empty_count = mas_mt_height(mas);
struct maple_subtree_state mast;
unsigned char shift, b_end = ++b_node->b_end;
MA_STATE(l_mas, mas->tree, mas->index, mas->last);
MA_STATE(r_mas, mas->tree, mas->index, mas->last);
trace_ma_op(__func__, mas);
/*
* Rebalancing occurs if a node is insufficient. Data is rebalanced
* against the node to the right if it exists, otherwise the node to the
* left of this node is rebalanced against this node. If rebalancing
* causes just one node to be produced instead of two, then the parent
* is also examined and rebalanced if it is insufficient. Every level
* tries to combine the data in the same way. If one node contains the
* entire range of the tree, then that node is used as a new root node.
*/
mas_node_count(mas, 1 + empty_count * 3);
if (mas_is_err(mas))
return 0;
mast.orig_l = &l_mas;
mast.orig_r = &r_mas;
mast.bn = b_node;
mast.bn->type = mte_node_type(mas->node);
l_mas = r_mas = *mas;
if (mas_next_sibling(&r_mas)) {
mas_mab_cp(&r_mas, 0, mt_slot_count(r_mas.node), b_node, b_end);
r_mas.last = r_mas.index = r_mas.max;
} else {
mas_prev_sibling(&l_mas);
shift = mas_data_end(&l_mas) + 1;
mab_shift_right(b_node, shift);
mas->offset += shift;
mas_mab_cp(&l_mas, 0, shift - 1, b_node, 0);
b_node->b_end = shift + b_end;
l_mas.index = l_mas.last = l_mas.min;
}
return mas_spanning_rebalance(mas, &mast, empty_count);
}
/*
* mas_destroy_rebalance() - Rebalance left-most node while destroying the maple
* state.
* @mas: The maple state
* @end: The end of the left-most node.
*
* During a mass-insert event (such as forking), it may be necessary to
* rebalance the left-most node when it is not sufficient.
*/
static inline void mas_destroy_rebalance(struct ma_state *mas, unsigned char end)
{
enum maple_type mt = mte_node_type(mas->node);
struct maple_node reuse, *newnode, *parent, *new_left, *left, *node;
struct maple_enode *eparent;
unsigned char offset, tmp, split = mt_slots[mt] / 2;
void __rcu **l_slots, **slots;
unsigned long *l_pivs, *pivs, gap;
bool in_rcu = mt_in_rcu(mas->tree);
MA_STATE(l_mas, mas->tree, mas->index, mas->last);
l_mas = *mas;
mas_prev_sibling(&l_mas);
/* set up node. */
if (in_rcu) {
/* Allocate for both left and right as well as parent. */
mas_node_count(mas, 3);
if (mas_is_err(mas))
return;
newnode = mas_pop_node(mas);
} else {
newnode = &reuse;
}
node = mas_mn(mas);
newnode->parent = node->parent;
slots = ma_slots(newnode, mt);
pivs = ma_pivots(newnode, mt);
left = mas_mn(&l_mas);
l_slots = ma_slots(left, mt);
l_pivs = ma_pivots(left, mt);
if (!l_slots[split])
split++;
tmp = mas_data_end(&l_mas) - split;
memcpy(slots, l_slots + split + 1, sizeof(void *) * tmp);
memcpy(pivs, l_pivs + split + 1, sizeof(unsigned long) * tmp);
pivs[tmp] = l_mas.max;
memcpy(slots + tmp, ma_slots(node, mt), sizeof(void *) * end);
memcpy(pivs + tmp, ma_pivots(node, mt), sizeof(unsigned long) * end);
l_mas.max = l_pivs[split];
mas->min = l_mas.max + 1;
eparent = mt_mk_node(mte_parent(l_mas.node),
mas_parent_enum(&l_mas, l_mas.node));
tmp += end;
if (!in_rcu) {
unsigned char max_p = mt_pivots[mt];
unsigned char max_s = mt_slots[mt];
if (tmp < max_p)
memset(pivs + tmp, 0,
sizeof(unsigned long *) * (max_p - tmp));
if (tmp < mt_slots[mt])
memset(slots + tmp, 0, sizeof(void *) * (max_s - tmp));
memcpy(node, newnode, sizeof(struct maple_node));
ma_set_meta(node, mt, 0, tmp - 1);
mte_set_pivot(eparent, mte_parent_slot(l_mas.node),
l_pivs[split]);
/* Remove data from l_pivs. */
tmp = split + 1;
memset(l_pivs + tmp, 0, sizeof(unsigned long) * (max_p - tmp));
memset(l_slots + tmp, 0, sizeof(void *) * (max_s - tmp));
ma_set_meta(left, mt, 0, split);
goto done;
}
/* RCU requires replacing both l_mas, mas, and parent. */
mas->node = mt_mk_node(newnode, mt);
ma_set_meta(newnode, mt, 0, tmp);
new_left = mas_pop_node(mas);
new_left->parent = left->parent;
mt = mte_node_type(l_mas.node);
slots = ma_slots(new_left, mt);
pivs = ma_pivots(new_left, mt);
memcpy(slots, l_slots, sizeof(void *) * split);
memcpy(pivs, l_pivs, sizeof(unsigned long) * split);
ma_set_meta(new_left, mt, 0, split);
l_mas.node = mt_mk_node(new_left, mt);
/* replace parent. */
offset = mte_parent_slot(mas->node);
mt = mas_parent_enum(&l_mas, l_mas.node);
parent = mas_pop_node(mas);
slots = ma_slots(parent, mt);
pivs = ma_pivots(parent, mt);
memcpy(parent, mte_to_node(eparent), sizeof(struct maple_node));
rcu_assign_pointer(slots[offset], mas->node);
rcu_assign_pointer(slots[offset - 1], l_mas.node);
pivs[offset - 1] = l_mas.max;
eparent = mt_mk_node(parent, mt);
done:
gap = mas_leaf_max_gap(mas);
mte_set_gap(eparent, mte_parent_slot(mas->node), gap);
gap = mas_leaf_max_gap(&l_mas);
mte_set_gap(eparent, mte_parent_slot(l_mas.node), gap);
mas_ascend(mas);
if (in_rcu)
mas_replace(mas, false);
mas_update_gap(mas);
}
/*
* mas_split_final_node() - Split the final node in a subtree operation.
* @mast: the maple subtree state
* @mas: The maple state
* @height: The height of the tree in case it's a new root.
*/
static inline bool mas_split_final_node(struct maple_subtree_state *mast,
struct ma_state *mas, int height)
{
struct maple_enode *ancestor;
if (mte_is_root(mas->node)) {
if (mt_is_alloc(mas->tree))
mast->bn->type = maple_arange_64;
else
mast->bn->type = maple_range_64;
mas->depth = height;
}
/*
* Only a single node is used here, could be root.
* The Big_node data should just fit in a single node.
*/
ancestor = mas_new_ma_node(mas, mast->bn);
mte_set_parent(mast->l->node, ancestor, mast->l->offset);
mte_set_parent(mast->r->node, ancestor, mast->r->offset);
mte_to_node(ancestor)->parent = mas_mn(mas)->parent;
mast->l->node = ancestor;
mab_mas_cp(mast->bn, 0, mt_slots[mast->bn->type] - 1, mast->l, true);
mas->offset = mast->bn->b_end - 1;
return true;
}
/*
* mast_fill_bnode() - Copy data into the big node in the subtree state
* @mast: The maple subtree state
* @mas: the maple state
* @skip: The number of entries to skip for new nodes insertion.
*/
static inline void mast_fill_bnode(struct maple_subtree_state *mast,
struct ma_state *mas,
unsigned char skip)
{
bool cp = true;
struct maple_enode *old = mas->node;
unsigned char split;
memset(mast->bn->gap, 0, sizeof(unsigned long) * ARRAY_SIZE(mast->bn->gap));
memset(mast->bn->slot, 0, sizeof(unsigned long) * ARRAY_SIZE(mast->bn->slot));
memset(mast->bn->pivot, 0, sizeof(unsigned long) * ARRAY_SIZE(mast->bn->pivot));
mast->bn->b_end = 0;
if (mte_is_root(mas->node)) {
cp = false;
} else {
mas_ascend(mas);
mat_add(mast->free, old);
mas->offset = mte_parent_slot(mas->node);
}
if (cp && mast->l->offset)
mas_mab_cp(mas, 0, mast->l->offset - 1, mast->bn, 0);
split = mast->bn->b_end;
mab_set_b_end(mast->bn, mast->l, mast->l->node);
mast->r->offset = mast->bn->b_end;
mab_set_b_end(mast->bn, mast->r, mast->r->node);
if (mast->bn->pivot[mast->bn->b_end - 1] == mas->max)
cp = false;
if (cp)
mas_mab_cp(mas, split + skip, mt_slot_count(mas->node) - 1,
mast->bn, mast->bn->b_end);
mast->bn->b_end--;
mast->bn->type = mte_node_type(mas->node);
}
/*
* mast_split_data() - Split the data in the subtree state big node into regular
* nodes.
* @mast: The maple subtree state
* @mas: The maple state
* @split: The location to split the big node
*/
static inline void mast_split_data(struct maple_subtree_state *mast,
struct ma_state *mas, unsigned char split)
{
unsigned char p_slot;
mab_mas_cp(mast->bn, 0, split, mast->l, true);
mte_set_pivot(mast->r->node, 0, mast->r->max);
mab_mas_cp(mast->bn, split + 1, mast->bn->b_end, mast->r, false);
mast->l->offset = mte_parent_slot(mas->node);
mast->l->max = mast->bn->pivot[split];
mast->r->min = mast->l->max + 1;
if (mte_is_leaf(mas->node))
return;
p_slot = mast->orig_l->offset;
mas_set_split_parent(mast->orig_l, mast->l->node, mast->r->node,
&p_slot, split);
mas_set_split_parent(mast->orig_r, mast->l->node, mast->r->node,
&p_slot, split);
}
/*
* mas_push_data() - Instead of splitting a node, it is beneficial to push the
* data to the right or left node if there is room.
* @mas: The maple state
* @height: The current height of the maple state
* @mast: The maple subtree state
* @left: Push left or not.
*
* Keeping the height of the tree low means faster lookups.
*
* Return: True if pushed, false otherwise.
*/
static inline bool mas_push_data(struct ma_state *mas, int height,
struct maple_subtree_state *mast, bool left)
{
unsigned char slot_total = mast->bn->b_end;
unsigned char end, space, split;
MA_STATE(tmp_mas, mas->tree, mas->index, mas->last);
tmp_mas = *mas;
tmp_mas.depth = mast->l->depth;
if (left && !mas_prev_sibling(&tmp_mas))
return false;
else if (!left && !mas_next_sibling(&tmp_mas))
return false;
end = mas_data_end(&tmp_mas);
slot_total += end;
space = 2 * mt_slot_count(mas->node) - 2;
/* -2 instead of -1 to ensure there isn't a triple split */
if (ma_is_leaf(mast->bn->type))
space--;
if (mas->max == ULONG_MAX)
space--;
if (slot_total >= space)
return false;
/* Get the data; Fill mast->bn */
mast->bn->b_end++;
if (left) {
mab_shift_right(mast->bn, end + 1);
mas_mab_cp(&tmp_mas, 0, end, mast->bn, 0);
mast->bn->b_end = slot_total + 1;
} else {
mas_mab_cp(&tmp_mas, 0, end, mast->bn, mast->bn->b_end);
}
/* Configure mast for splitting of mast->bn */
split = mt_slots[mast->bn->type] - 2;
if (left) {
/* Switch mas to prev node */
mat_add(mast->free, mas->node);
*mas = tmp_mas;
/* Start using mast->l for the left side. */
tmp_mas.node = mast->l->node;
*mast->l = tmp_mas;
} else {
mat_add(mast->free, tmp_mas.node);
tmp_mas.node = mast->r->node;
*mast->r = tmp_mas;
split = slot_total - split;
}
split = mab_no_null_split(mast->bn, split, mt_slots[mast->bn->type]);
/* Update parent slot for split calculation. */
if (left)
mast->orig_l->offset += end + 1;
mast_split_data(mast, mas, split);
mast_fill_bnode(mast, mas, 2);
mas_split_final_node(mast, mas, height + 1);
return true;
}
/*
* mas_split() - Split data that is too big for one node into two.
* @mas: The maple state
* @b_node: The maple big node
* Return: 1 on success, 0 on failure.
*/
static int mas_split(struct ma_state *mas, struct maple_big_node *b_node)
{
struct maple_subtree_state mast;
int height = 0;
unsigned char mid_split, split = 0;
/*
* Splitting is handled differently from any other B-tree; the Maple
* Tree splits upwards. Splitting up means that the split operation
* occurs when the walk of the tree hits the leaves and not on the way
* down. The reason for splitting up is that it is impossible to know
* how much space will be needed until the leaf is (or leaves are)
* reached. Since overwriting data is allowed and a range could
* overwrite more than one range or result in changing one entry into 3
* entries, it is impossible to know if a split is required until the
* data is examined.
*
* Splitting is a balancing act between keeping allocations to a minimum
* and avoiding a 'jitter' event where a tree is expanded to make room
* for an entry followed by a contraction when the entry is removed. To
* accomplish the balance, there are empty slots remaining in both left
* and right nodes after a split.
*/
MA_STATE(l_mas, mas->tree, mas->index, mas->last);
MA_STATE(r_mas, mas->tree, mas->index, mas->last);
MA_STATE(prev_l_mas, mas->tree, mas->index, mas->last);
MA_STATE(prev_r_mas, mas->tree, mas->index, mas->last);
MA_TOPIARY(mat, mas->tree);
trace_ma_op(__func__, mas);
mas->depth = mas_mt_height(mas);
/* Allocation failures will happen early. */
mas_node_count(mas, 1 + mas->depth * 2);
if (mas_is_err(mas))
return 0;
mast.l = &l_mas;
mast.r = &r_mas;
mast.orig_l = &prev_l_mas;
mast.orig_r = &prev_r_mas;
mast.free = &mat;
mast.bn = b_node;
while (height++ <= mas->depth) {
if (mt_slots[b_node->type] > b_node->b_end) {
mas_split_final_node(&mast, mas, height);
break;
}
l_mas = r_mas = *mas;
l_mas.node = mas_new_ma_node(mas, b_node);
r_mas.node = mas_new_ma_node(mas, b_node);
/*
* Another way that 'jitter' is avoided is to terminate a split up early if the
* left or right node has space to spare. This is referred to as "pushing left"
* or "pushing right" and is similar to the B* tree, except the nodes left or
* right can rarely be reused due to RCU, but the ripple upwards is halted which
* is a significant savings.
*/
/* Try to push left. */
if (mas_push_data(mas, height, &mast, true))
break;
/* Try to push right. */
if (mas_push_data(mas, height, &mast, false))
break;
split = mab_calc_split(mas, b_node, &mid_split, prev_l_mas.min);
mast_split_data(&mast, mas, split);
/*
* Usually correct, mab_mas_cp in the above call overwrites
* r->max.
*/
mast.r->max = mas->max;
mast_fill_bnode(&mast, mas, 1);
prev_l_mas = *mast.l;
prev_r_mas = *mast.r;
}
/* Set the original node as dead */
mat_add(mast.free, mas->node);
mas->node = l_mas.node;
mas_wmb_replace(mas, mast.free, NULL);
mtree_range_walk(mas);
return 1;
}
/*
* mas_reuse_node() - Reuse the node to store the data.
* @wr_mas: The maple write state
* @bn: The maple big node
* @end: The end of the data.
*
* Will always return false in RCU mode.
*
* Return: True if node was reused, false otherwise.
*/
static inline bool mas_reuse_node(struct ma_wr_state *wr_mas,
struct maple_big_node *bn, unsigned char end)
{
/* Need to be rcu safe. */
if (mt_in_rcu(wr_mas->mas->tree))
return false;
if (end > bn->b_end) {
int clear = mt_slots[wr_mas->type] - bn->b_end;
memset(wr_mas->slots + bn->b_end, 0, sizeof(void *) * clear--);
memset(wr_mas->pivots + bn->b_end, 0, sizeof(void *) * clear);
}
mab_mas_cp(bn, 0, bn->b_end, wr_mas->mas, false);
return true;
}
/*
* mas_commit_b_node() - Commit the big node into the tree.
* @wr_mas: The maple write state
* @b_node: The maple big node
* @end: The end of the data.
*/
static noinline_for_kasan int mas_commit_b_node(struct ma_wr_state *wr_mas,
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
struct maple_big_node *b_node, unsigned char end)
{
struct maple_node *node;
unsigned char b_end = b_node->b_end;
enum maple_type b_type = b_node->type;
if ((b_end < mt_min_slots[b_type]) &&
(!mte_is_root(wr_mas->mas->node)) &&
(mas_mt_height(wr_mas->mas) > 1))
return mas_rebalance(wr_mas->mas, b_node);
if (b_end >= mt_slots[b_type])
return mas_split(wr_mas->mas, b_node);
if (mas_reuse_node(wr_mas, b_node, end))
goto reuse_node;
mas_node_count(wr_mas->mas, 1);
if (mas_is_err(wr_mas->mas))
return 0;
node = mas_pop_node(wr_mas->mas);
node->parent = mas_mn(wr_mas->mas)->parent;
wr_mas->mas->node = mt_mk_node(node, b_type);
mab_mas_cp(b_node, 0, b_end, wr_mas->mas, false);
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
mas_replace(wr_mas->mas, false);
reuse_node:
mas_update_gap(wr_mas->mas);
return 1;
}
/*
* mas_root_expand() - Expand a root to a node
* @mas: The maple state
* @entry: The entry to store into the tree
*/
static inline int mas_root_expand(struct ma_state *mas, void *entry)
{
void *contents = mas_root_locked(mas);
enum maple_type type = maple_leaf_64;
struct maple_node *node;
void __rcu **slots;
unsigned long *pivots;
int slot = 0;
mas_node_count(mas, 1);
if (unlikely(mas_is_err(mas)))
return 0;
node = mas_pop_node(mas);
pivots = ma_pivots(node, type);
slots = ma_slots(node, type);
node->parent = ma_parent_ptr(
((unsigned long)mas->tree | MA_ROOT_PARENT));
mas->node = mt_mk_node(node, type);
if (mas->index) {
if (contents) {
rcu_assign_pointer(slots[slot], contents);
if (likely(mas->index > 1))
slot++;
}
pivots[slot++] = mas->index - 1;
}
rcu_assign_pointer(slots[slot], entry);
mas->offset = slot;
pivots[slot] = mas->last;
if (mas->last != ULONG_MAX)
slot++;
mas->depth = 1;
mas_set_height(mas);
/* swap the new root into the tree */
rcu_assign_pointer(mas->tree->ma_root, mte_mk_root(mas->node));
ma_set_meta(node, maple_leaf_64, 0, slot);
return slot;
}
static inline void mas_store_root(struct ma_state *mas, void *entry)
{
if (likely((mas->last != 0) || (mas->index != 0)))
mas_root_expand(mas, entry);
else if (((unsigned long) (entry) & 3) == 2)
mas_root_expand(mas, entry);
else {
rcu_assign_pointer(mas->tree->ma_root, entry);
mas->node = MAS_START;
}
}
/*
* mas_is_span_wr() - Check if the write needs to be treated as a write that
* spans the node.
* @mas: The maple state
* @piv: The pivot value being written
* @type: The maple node type
* @entry: The data to write
*
* Spanning writes are writes that start in one node and end in another OR if
* the write of a %NULL will cause the node to end with a %NULL.
*
* Return: True if this is a spanning write, false otherwise.
*/
static bool mas_is_span_wr(struct ma_wr_state *wr_mas)
{
unsigned long max;
unsigned long last = wr_mas->mas->last;
unsigned long piv = wr_mas->r_max;
enum maple_type type = wr_mas->type;
void *entry = wr_mas->entry;
/* Contained in this pivot */
if (piv > last)
return false;
max = wr_mas->mas->max;
if (unlikely(ma_is_leaf(type))) {
/* Fits in the node, but may span slots. */
if (last < max)
return false;
/* Writes to the end of the node but not null. */
if ((last == max) && entry)
return false;
/*
* Writing ULONG_MAX is not a spanning write regardless of the
* value being written as long as the range fits in the node.
*/
if ((last == ULONG_MAX) && (last == max))
return false;
} else if (piv == last) {
if (entry)
return false;
/* Detect spanning store wr walk */
if (last == ULONG_MAX)
return false;
}
trace_ma_write(__func__, wr_mas->mas, piv, entry);
return true;
}
static inline void mas_wr_walk_descend(struct ma_wr_state *wr_mas)
{
wr_mas->type = mte_node_type(wr_mas->mas->node);
mas_wr_node_walk(wr_mas);
wr_mas->slots = ma_slots(wr_mas->node, wr_mas->type);
}
static inline void mas_wr_walk_traverse(struct ma_wr_state *wr_mas)
{
wr_mas->mas->max = wr_mas->r_max;
wr_mas->mas->min = wr_mas->r_min;
wr_mas->mas->node = wr_mas->content;
wr_mas->mas->offset = 0;
wr_mas->mas->depth++;
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
}
/*
* mas_wr_walk() - Walk the tree for a write.
* @wr_mas: The maple write state
*
* Uses mas_slot_locked() and does not need to worry about dead nodes.
*
* Return: True if it's contained in a node, false on spanning write.
*/
static bool mas_wr_walk(struct ma_wr_state *wr_mas)
{
struct ma_state *mas = wr_mas->mas;
while (true) {
mas_wr_walk_descend(wr_mas);
if (unlikely(mas_is_span_wr(wr_mas)))
return false;
wr_mas->content = mas_slot_locked(mas, wr_mas->slots,
mas->offset);
if (ma_is_leaf(wr_mas->type))
return true;
mas_wr_walk_traverse(wr_mas);
}
return true;
}
static bool mas_wr_walk_index(struct ma_wr_state *wr_mas)
{
struct ma_state *mas = wr_mas->mas;
while (true) {
mas_wr_walk_descend(wr_mas);
wr_mas->content = mas_slot_locked(mas, wr_mas->slots,
mas->offset);
if (ma_is_leaf(wr_mas->type))
return true;
mas_wr_walk_traverse(wr_mas);
}
return true;
}
/*
* mas_extend_spanning_null() - Extend a store of a %NULL to include surrounding %NULLs.
* @l_wr_mas: The left maple write state
* @r_wr_mas: The right maple write state
*/
static inline void mas_extend_spanning_null(struct ma_wr_state *l_wr_mas,
struct ma_wr_state *r_wr_mas)
{
struct ma_state *r_mas = r_wr_mas->mas;
struct ma_state *l_mas = l_wr_mas->mas;
unsigned char l_slot;
l_slot = l_mas->offset;
if (!l_wr_mas->content)
l_mas->index = l_wr_mas->r_min;
if ((l_mas->index == l_wr_mas->r_min) &&
(l_slot &&
!mas_slot_locked(l_mas, l_wr_mas->slots, l_slot - 1))) {
if (l_slot > 1)
l_mas->index = l_wr_mas->pivots[l_slot - 2] + 1;
else
l_mas->index = l_mas->min;
l_mas->offset = l_slot - 1;
}
if (!r_wr_mas->content) {
if (r_mas->last < r_wr_mas->r_max)
r_mas->last = r_wr_mas->r_max;
r_mas->offset++;
} else if ((r_mas->last == r_wr_mas->r_max) &&
(r_mas->last < r_mas->max) &&
!mas_slot_locked(r_mas, r_wr_mas->slots, r_mas->offset + 1)) {
r_mas->last = mas_safe_pivot(r_mas, r_wr_mas->pivots,
r_wr_mas->type, r_mas->offset + 1);
r_mas->offset++;
}
}
static inline void *mas_state_walk(struct ma_state *mas)
{
void *entry;
entry = mas_start(mas);
if (mas_is_none(mas))
return NULL;
if (mas_is_ptr(mas))
return entry;
return mtree_range_walk(mas);
}
/*
* mtree_lookup_walk() - Internal quick lookup that does not keep maple state up
* to date.
*
* @mas: The maple state.
*
* Note: Leaves mas in undesirable state.
* Return: The entry for @mas->index or %NULL on dead node.
*/
static inline void *mtree_lookup_walk(struct ma_state *mas)
{
unsigned long *pivots;
unsigned char offset;
struct maple_node *node;
struct maple_enode *next;
enum maple_type type;
void __rcu **slots;
unsigned char end;
unsigned long max;
next = mas->node;
max = ULONG_MAX;
do {
offset = 0;
node = mte_to_node(next);
type = mte_node_type(next);
pivots = ma_pivots(node, type);
end = ma_data_end(node, type, pivots, max);
if (unlikely(ma_dead_node(node)))
goto dead_node;
if (pivots[offset] >= mas->index)
goto next;
do {
offset++;
} while ((offset < end) && (pivots[offset] < mas->index));
if (likely(offset > end))
max = pivots[offset];
next:
slots = ma_slots(node, type);
next = mt_slot(mas->tree, slots, offset);
if (unlikely(ma_dead_node(node)))
goto dead_node;
} while (!ma_is_leaf(type));
return (void *)next;
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
dead_node:
mas_reset(mas);
return NULL;
}
/*
* mas_new_root() - Create a new root node that only contains the entry passed
* in.
* @mas: The maple state
* @entry: The entry to store.
*
* Only valid when the index == 0 and the last == ULONG_MAX
*
* Return 0 on error, 1 on success.
*/
static inline int mas_new_root(struct ma_state *mas, void *entry)
{
struct maple_enode *root = mas_root_locked(mas);
enum maple_type type = maple_leaf_64;
struct maple_node *node;
void __rcu **slots;
unsigned long *pivots;
if (!entry && !mas->index && mas->last == ULONG_MAX) {
mas->depth = 0;
mas_set_height(mas);
rcu_assign_pointer(mas->tree->ma_root, entry);
mas->node = MAS_START;
goto done;
}
mas_node_count(mas, 1);
if (mas_is_err(mas))
return 0;
node = mas_pop_node(mas);
pivots = ma_pivots(node, type);
slots = ma_slots(node, type);
node->parent = ma_parent_ptr(
((unsigned long)mas->tree | MA_ROOT_PARENT));
mas->node = mt_mk_node(node, type);
rcu_assign_pointer(slots[0], entry);
pivots[0] = mas->last;
mas->depth = 1;
mas_set_height(mas);
rcu_assign_pointer(mas->tree->ma_root, mte_mk_root(mas->node));
done:
if (xa_is_node(root))
mte_destroy_walk(root, mas->tree);
return 1;
}
/*
* mas_wr_spanning_store() - Create a subtree with the store operation completed
* and new nodes where necessary, then place the sub-tree in the actual tree.
* Note that mas is expected to point to the node which caused the store to
* span.
* @wr_mas: The maple write state
*
* Return: 0 on error, positive on success.
*/
static inline int mas_wr_spanning_store(struct ma_wr_state *wr_mas)
{
struct maple_subtree_state mast;
struct maple_big_node b_node;
struct ma_state *mas;
unsigned char height;
/* Left and Right side of spanning store */
MA_STATE(l_mas, NULL, 0, 0);
MA_STATE(r_mas, NULL, 0, 0);
MA_WR_STATE(r_wr_mas, &r_mas, wr_mas->entry);
MA_WR_STATE(l_wr_mas, &l_mas, wr_mas->entry);
/*
* A store operation that spans multiple nodes is called a spanning
* store and is handled early in the store call stack by the function
* mas_is_span_wr(). When a spanning store is identified, the maple
* state is duplicated. The first maple state walks the left tree path
* to ``index``, the duplicate walks the right tree path to ``last``.
* The data in the two nodes are combined into a single node, two nodes,
* or possibly three nodes (see the 3-way split above). A ``NULL``
* written to the last entry of a node is considered a spanning store as
* a rebalance is required for the operation to complete and an overflow
* of data may happen.
*/
mas = wr_mas->mas;
trace_ma_op(__func__, mas);
if (unlikely(!mas->index && mas->last == ULONG_MAX))
return mas_new_root(mas, wr_mas->entry);
/*
* Node rebalancing may occur due to this store, so there may be three new
* entries per level plus a new root.
*/
height = mas_mt_height(mas);
mas_node_count(mas, 1 + height * 3);
if (mas_is_err(mas))
return 0;
/*
* Set up right side. Need to get to the next offset after the spanning
* store to ensure it's not NULL and to combine both the next node and
* the node with the start together.
*/
r_mas = *mas;
/* Avoid overflow, walk to next slot in the tree. */
if (r_mas.last + 1)
r_mas.last++;
r_mas.index = r_mas.last;
mas_wr_walk_index(&r_wr_mas);
r_mas.last = r_mas.index = mas->last;
/* Set up left side. */
l_mas = *mas;
mas_wr_walk_index(&l_wr_mas);
if (!wr_mas->entry) {
mas_extend_spanning_null(&l_wr_mas, &r_wr_mas);
mas->offset = l_mas.offset;
mas->index = l_mas.index;
mas->last = l_mas.last = r_mas.last;
}
/* expanding NULLs may make this cover the entire range */
if (!l_mas.index && r_mas.last == ULONG_MAX) {
mas_set_range(mas, 0, ULONG_MAX);
return mas_new_root(mas, wr_mas->entry);
}
memset(&b_node, 0, sizeof(struct maple_big_node));
/* Copy l_mas and store the value in b_node. */
mas_store_b_node(&l_wr_mas, &b_node, l_wr_mas.node_end);
/* Copy r_mas into b_node. */
if (r_mas.offset <= r_wr_mas.node_end)
mas_mab_cp(&r_mas, r_mas.offset, r_wr_mas.node_end,
&b_node, b_node.b_end + 1);
else
b_node.b_end++;
/* Stop spanning searches by searching for just index. */
l_mas.index = l_mas.last = mas->index;
mast.bn = &b_node;
mast.orig_l = &l_mas;
mast.orig_r = &r_mas;
/* Combine l_mas and r_mas and split them up evenly again. */
return mas_spanning_rebalance(mas, &mast, height + 1);
}
/*
* mas_wr_node_store() - Attempt to store the value in a node
* @wr_mas: The maple write state
*
* Attempts to reuse the node, but may allocate.
*
* Return: True if stored, false otherwise
*/
static inline bool mas_wr_node_store(struct ma_wr_state *wr_mas)
{
struct ma_state *mas = wr_mas->mas;
void __rcu **dst_slots;
unsigned long *dst_pivots;
unsigned char dst_offset;
unsigned char new_end = wr_mas->node_end;
unsigned char offset;
unsigned char node_slots = mt_slots[wr_mas->type];
struct maple_node reuse, *newnode;
unsigned char copy_size, max_piv = mt_pivots[wr_mas->type];
bool in_rcu = mt_in_rcu(mas->tree);
offset = mas->offset;
if (mas->last == wr_mas->r_max) {
/* runs right to the end of the node */
if (mas->last == mas->max)
new_end = offset;
/* don't copy this offset */
wr_mas->offset_end++;
} else if (mas->last < wr_mas->r_max) {
/* new range ends in this range */
if (unlikely(wr_mas->r_max == ULONG_MAX))
mas_bulk_rebalance(mas, wr_mas->node_end, wr_mas->type);
new_end++;
} else {
if (wr_mas->end_piv == mas->last)
wr_mas->offset_end++;
new_end -= wr_mas->offset_end - offset - 1;
}
/* new range starts within a range */
if (wr_mas->r_min < mas->index)
new_end++;
/* Not enough room */
if (new_end >= node_slots)
return false;
/* Not enough data. */
if (!mte_is_root(mas->node) && (new_end <= mt_min_slots[wr_mas->type]) &&
!(mas->mas_flags & MA_STATE_BULK))
return false;
/* set up node. */
if (in_rcu) {
mas_node_count(mas, 1);
if (mas_is_err(mas))
return false;
newnode = mas_pop_node(mas);
} else {
memset(&reuse, 0, sizeof(struct maple_node));
newnode = &reuse;
}
newnode->parent = mas_mn(mas)->parent;
dst_pivots = ma_pivots(newnode, wr_mas->type);
dst_slots = ma_slots(newnode, wr_mas->type);
/* Copy from start to insert point */
memcpy(dst_pivots, wr_mas->pivots, sizeof(unsigned long) * (offset + 1));
memcpy(dst_slots, wr_mas->slots, sizeof(void *) * (offset + 1));
dst_offset = offset;
/* Handle insert of new range starting after old range */
if (wr_mas->r_min < mas->index) {
mas->offset++;
rcu_assign_pointer(dst_slots[dst_offset], wr_mas->content);
dst_pivots[dst_offset++] = mas->index - 1;
}
/* Store the new entry and range end. */
if (dst_offset < max_piv)
dst_pivots[dst_offset] = mas->last;
mas->offset = dst_offset;
rcu_assign_pointer(dst_slots[dst_offset], wr_mas->entry);
/*
* this range wrote to the end of the node or it overwrote the rest of
* the data
*/
if (wr_mas->offset_end > wr_mas->node_end || mas->last >= mas->max) {
new_end = dst_offset;
goto done;
}
dst_offset++;
/* Copy to the end of node if necessary. */
copy_size = wr_mas->node_end - wr_mas->offset_end + 1;
memcpy(dst_slots + dst_offset, wr_mas->slots + wr_mas->offset_end,
sizeof(void *) * copy_size);
if (dst_offset < max_piv) {
if (copy_size > max_piv - dst_offset)
copy_size = max_piv - dst_offset;
memcpy(dst_pivots + dst_offset,
wr_mas->pivots + wr_mas->offset_end,
sizeof(unsigned long) * copy_size);
}
if ((wr_mas->node_end == node_slots - 1) && (new_end < node_slots - 1))
dst_pivots[new_end] = mas->max;
done:
mas_leaf_set_meta(mas, newnode, dst_pivots, maple_leaf_64, new_end);
if (in_rcu) {
mas->node = mt_mk_node(newnode, wr_mas->type);
mas_replace(mas, false);
} else {
memcpy(wr_mas->node, newnode, sizeof(struct maple_node));
}
trace_ma_write(__func__, mas, 0, wr_mas->entry);
mas_update_gap(mas);
return true;
}
/*
* mas_wr_slot_store: Attempt to store a value in a slot.
* @wr_mas: the maple write state
*
* Return: True if stored, false otherwise
*/
static inline bool mas_wr_slot_store(struct ma_wr_state *wr_mas)
{
struct ma_state *mas = wr_mas->mas;
unsigned long lmax; /* Logical max. */
unsigned char offset = mas->offset;
if ((wr_mas->r_max > mas->last) && ((wr_mas->r_min != mas->index) ||
(offset != wr_mas->node_end)))
return false;
if (offset == wr_mas->node_end - 1)
lmax = mas->max;
else
lmax = wr_mas->pivots[offset + 1];
/* going to overwrite too many slots. */
if (lmax < mas->last)
return false;
if (wr_mas->r_min == mas->index) {
/* overwriting two or more ranges with one. */
if (lmax == mas->last)
return false;
/* Overwriting all of offset and a portion of offset + 1. */
rcu_assign_pointer(wr_mas->slots[offset], wr_mas->entry);
wr_mas->pivots[offset] = mas->last;
goto done;
}
/* Doesn't end on the next range end. */
if (lmax != mas->last)
return false;
/* Overwriting a portion of offset and all of offset + 1 */
if ((offset + 1 < mt_pivots[wr_mas->type]) &&
(wr_mas->entry || wr_mas->pivots[offset + 1]))
wr_mas->pivots[offset + 1] = mas->last;
rcu_assign_pointer(wr_mas->slots[offset + 1], wr_mas->entry);
wr_mas->pivots[offset] = mas->index - 1;
mas->offset++; /* Keep mas accurate. */
done:
trace_ma_write(__func__, mas, 0, wr_mas->entry);
mas_update_gap(mas);
return true;
}
static inline void mas_wr_end_piv(struct ma_wr_state *wr_mas)
{
while ((wr_mas->mas->last > wr_mas->end_piv) &&
(wr_mas->offset_end < wr_mas->node_end))
wr_mas->end_piv = wr_mas->pivots[++wr_mas->offset_end];
if (wr_mas->mas->last > wr_mas->end_piv)
wr_mas->end_piv = wr_mas->mas->max;
}
static inline void mas_wr_extend_null(struct ma_wr_state *wr_mas)
{
struct ma_state *mas = wr_mas->mas;
if (mas->last < wr_mas->end_piv && !wr_mas->slots[wr_mas->offset_end])
mas->last = wr_mas->end_piv;
/* Check next slot(s) if we are overwriting the end */
if ((mas->last == wr_mas->end_piv) &&
(wr_mas->node_end != wr_mas->offset_end) &&
!wr_mas->slots[wr_mas->offset_end + 1]) {
wr_mas->offset_end++;
if (wr_mas->offset_end == wr_mas->node_end)
mas->last = mas->max;
else
mas->last = wr_mas->pivots[wr_mas->offset_end];
wr_mas->end_piv = mas->last;
}
if (!wr_mas->content) {
/* If this one is null, the next and prev are not */
mas->index = wr_mas->r_min;
} else {
/* Check prev slot if we are overwriting the start */
if (mas->index == wr_mas->r_min && mas->offset &&
!wr_mas->slots[mas->offset - 1]) {
mas->offset--;
wr_mas->r_min = mas->index =
mas_safe_min(mas, wr_mas->pivots, mas->offset);
wr_mas->r_max = wr_mas->pivots[mas->offset];
}
}
}
static inline bool mas_wr_append(struct ma_wr_state *wr_mas)
{
unsigned char end = wr_mas->node_end;
unsigned char new_end = end + 1;
struct ma_state *mas = wr_mas->mas;
unsigned char node_pivots = mt_pivots[wr_mas->type];
if ((mas->index != wr_mas->r_min) && (mas->last == wr_mas->r_max)) {
if (new_end < node_pivots)
wr_mas->pivots[new_end] = wr_mas->pivots[end];
if (new_end < node_pivots)
ma_set_meta(wr_mas->node, maple_leaf_64, 0, new_end);
rcu_assign_pointer(wr_mas->slots[new_end], wr_mas->entry);
mas->offset = new_end;
wr_mas->pivots[end] = mas->index - 1;
return true;
}
if ((mas->index == wr_mas->r_min) && (mas->last < wr_mas->r_max)) {
if (new_end < node_pivots)
wr_mas->pivots[new_end] = wr_mas->pivots[end];
rcu_assign_pointer(wr_mas->slots[new_end], wr_mas->content);
if (new_end < node_pivots)
ma_set_meta(wr_mas->node, maple_leaf_64, 0, new_end);
wr_mas->pivots[end] = mas->last;
rcu_assign_pointer(wr_mas->slots[end], wr_mas->entry);
return true;
}
return false;
}
/*
* mas_wr_bnode() - Slow path for a modification.
* @wr_mas: The write maple state
*
* This is where split, rebalance end up.
*/
static void mas_wr_bnode(struct ma_wr_state *wr_mas)
{
struct maple_big_node b_node;
trace_ma_write(__func__, wr_mas->mas, 0, wr_mas->entry);
memset(&b_node, 0, sizeof(struct maple_big_node));
mas_store_b_node(wr_mas, &b_node, wr_mas->offset_end);
mas_commit_b_node(wr_mas, &b_node, wr_mas->node_end);
}
static inline void mas_wr_modify(struct ma_wr_state *wr_mas)
{
unsigned char node_slots;
unsigned char node_size;
struct ma_state *mas = wr_mas->mas;
/* Direct replacement */
if (wr_mas->r_min == mas->index && wr_mas->r_max == mas->last) {
rcu_assign_pointer(wr_mas->slots[mas->offset], wr_mas->entry);
if (!!wr_mas->entry ^ !!wr_mas->content)
mas_update_gap(mas);
return;
}
/* Attempt to append */
node_slots = mt_slots[wr_mas->type];
node_size = wr_mas->node_end - wr_mas->offset_end + mas->offset + 2;
if (mas->max == ULONG_MAX)
node_size++;
/* slot and node store will not fit, go to the slow path */
if (unlikely(node_size >= node_slots))
goto slow_path;
if (wr_mas->entry && (wr_mas->node_end < node_slots - 1) &&
(mas->offset == wr_mas->node_end) && mas_wr_append(wr_mas)) {
if (!wr_mas->content || !wr_mas->entry)
mas_update_gap(mas);
return;
}
if ((wr_mas->offset_end - mas->offset <= 1) && mas_wr_slot_store(wr_mas))
return;
else if (mas_wr_node_store(wr_mas))
return;
if (mas_is_err(mas))
return;
slow_path:
mas_wr_bnode(wr_mas);
}
/*
* mas_wr_store_entry() - Internal call to store a value
* @mas: The maple state
* @entry: The entry to store.
*
* Return: The contents that was stored at the index.
*/
static inline void *mas_wr_store_entry(struct ma_wr_state *wr_mas)
{
struct ma_state *mas = wr_mas->mas;
wr_mas->content = mas_start(mas);
if (mas_is_none(mas) || mas_is_ptr(mas)) {
mas_store_root(mas, wr_mas->entry);
return wr_mas->content;
}
if (unlikely(!mas_wr_walk(wr_mas))) {
mas_wr_spanning_store(wr_mas);
return wr_mas->content;
}
/* At this point, we are at the leaf node that needs to be altered. */
wr_mas->end_piv = wr_mas->r_max;
mas_wr_end_piv(wr_mas);
if (!wr_mas->entry)
mas_wr_extend_null(wr_mas);
/* New root for a single pointer */
if (unlikely(!mas->index && mas->last == ULONG_MAX)) {
mas_new_root(mas, wr_mas->entry);
return wr_mas->content;
}
mas_wr_modify(wr_mas);
return wr_mas->content;
}
/**
* mas_insert() - Internal call to insert a value
* @mas: The maple state
* @entry: The entry to store
*
* Return: %NULL or the contents that already exists at the requested index
* otherwise. The maple state needs to be checked for error conditions.
*/
static inline void *mas_insert(struct ma_state *mas, void *entry)
{
MA_WR_STATE(wr_mas, mas, entry);
/*
* Inserting a new range inserts either 0, 1, or 2 pivots within the
* tree. If the insert fits exactly into an existing gap with a value
* of NULL, then the slot only needs to be written with the new value.
* If the range being inserted is adjacent to another range, then only a
* single pivot needs to be inserted (as well as writing the entry). If
* the new range is within a gap but does not touch any other ranges,
* then two pivots need to be inserted: the start - 1, and the end. As
* usual, the entry must be written. Most operations require a new node
* to be allocated and replace an existing node to ensure RCU safety,
* when in RCU mode. The exception to requiring a newly allocated node
* is when inserting at the end of a node (appending). When done
* carefully, appending can reuse the node in place.
*/
wr_mas.content = mas_start(mas);
if (wr_mas.content)
goto exists;
if (mas_is_none(mas) || mas_is_ptr(mas)) {
mas_store_root(mas, entry);
return NULL;
}
/* spanning writes always overwrite something */
if (!mas_wr_walk(&wr_mas))
goto exists;
/* At this point, we are at the leaf node that needs to be altered. */
wr_mas.offset_end = mas->offset;
wr_mas.end_piv = wr_mas.r_max;
if (wr_mas.content || (mas->last > wr_mas.r_max))
goto exists;
if (!entry)
return NULL;
mas_wr_modify(&wr_mas);
return wr_mas.content;
exists:
mas_set_err(mas, -EEXIST);
return wr_mas.content;
}
/*
* mas_prev_node() - Find the prev non-null entry at the same level in the
* tree. The prev value will be mas->node[mas->offset] or MAS_NONE.
* @mas: The maple state
* @min: The lower limit to search
*
* The prev node value will be mas->node[mas->offset] or MAS_NONE.
* Return: 1 if the node is dead, 0 otherwise.
*/
static inline int mas_prev_node(struct ma_state *mas, unsigned long min)
{
enum maple_type mt;
int offset, level;
void __rcu **slots;
struct maple_node *node;
struct maple_enode *enode;
unsigned long *pivots;
if (mas_is_none(mas))
return 0;
level = 0;
do {
node = mas_mn(mas);
if (ma_is_root(node))
goto no_entry;
/* Walk up. */
if (unlikely(mas_ascend(mas)))
return 1;
offset = mas->offset;
level++;
} while (!offset);
offset--;
mt = mte_node_type(mas->node);
node = mas_mn(mas);
slots = ma_slots(node, mt);
pivots = ma_pivots(node, mt);
mas->max = pivots[offset];
if (offset)
mas->min = pivots[offset - 1] + 1;
if (unlikely(ma_dead_node(node)))
return 1;
if (mas->max < min)
goto no_entry_min;
while (level > 1) {
level--;
enode = mas_slot(mas, slots, offset);
if (unlikely(ma_dead_node(node)))
return 1;
mas->node = enode;
mt = mte_node_type(mas->node);
node = mas_mn(mas);
slots = ma_slots(node, mt);
pivots = ma_pivots(node, mt);
offset = ma_data_end(node, mt, pivots, mas->max);
if (offset)
mas->min = pivots[offset - 1] + 1;
if (offset < mt_pivots[mt])
mas->max = pivots[offset];
if (mas->max < min)
goto no_entry;
}
mas->node = mas_slot(mas, slots, offset);
if (unlikely(ma_dead_node(node)))
return 1;
mas->offset = mas_data_end(mas);
if (unlikely(mte_dead_node(mas->node)))
return 1;
return 0;
no_entry_min:
mas->offset = offset;
if (offset)
mas->min = pivots[offset - 1] + 1;
no_entry:
if (unlikely(ma_dead_node(node)))
return 1;
mas->node = MAS_NONE;
return 0;
}
/*
* mas_next_node() - Get the next node at the same level in the tree.
* @mas: The maple state
* @max: The maximum pivot value to check.
*
* The next value will be mas->node[mas->offset] or MAS_NONE.
* Return: 1 on dead node, 0 otherwise.
*/
static inline int mas_next_node(struct ma_state *mas, struct maple_node *node,
unsigned long max)
{
unsigned long min, pivot;
unsigned long *pivots;
struct maple_enode *enode;
int level = 0;
unsigned char offset;
enum maple_type mt;
void __rcu **slots;
if (mas->max >= max)
goto no_entry;
level = 0;
do {
if (ma_is_root(node))
goto no_entry;
min = mas->max + 1;
if (min > max)
goto no_entry;
if (unlikely(mas_ascend(mas)))
return 1;
offset = mas->offset;
level++;
node = mas_mn(mas);
mt = mte_node_type(mas->node);
pivots = ma_pivots(node, mt);
} while (unlikely(offset == ma_data_end(node, mt, pivots, mas->max)));
slots = ma_slots(node, mt);
pivot = mas_safe_pivot(mas, pivots, ++offset, mt);
while (unlikely(level > 1)) {
/* Descend, if necessary */
enode = mas_slot(mas, slots, offset);
if (unlikely(ma_dead_node(node)))
return 1;
mas->node = enode;
level--;
node = mas_mn(mas);
mt = mte_node_type(mas->node);
slots = ma_slots(node, mt);
pivots = ma_pivots(node, mt);
offset = 0;
pivot = pivots[0];
}
enode = mas_slot(mas, slots, offset);
if (unlikely(ma_dead_node(node)))
return 1;
mas->node = enode;
mas->min = min;
mas->max = pivot;
return 0;
no_entry:
if (unlikely(ma_dead_node(node)))
return 1;
mas->node = MAS_NONE;
return 0;
}
/*
* mas_next_nentry() - Get the next node entry
* @mas: The maple state
* @max: The maximum value to check
* @*range_start: Pointer to store the start of the range.
*
* Sets @mas->offset to the offset of the next node entry, @mas->last to the
* pivot of the entry.
*
* Return: The next entry, %NULL otherwise
*/
static inline void *mas_next_nentry(struct ma_state *mas,
struct maple_node *node, unsigned long max, enum maple_type type)
{
unsigned char count;
unsigned long pivot;
unsigned long *pivots;
void __rcu **slots;
void *entry;
if (mas->last == mas->max) {
mas->index = mas->max;
return NULL;
}
pivots = ma_pivots(node, type);
slots = ma_slots(node, type);
mas->index = mas_safe_min(mas, pivots, mas->offset);
count = ma_data_end(node, type, pivots, mas->max);
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
if (ma_dead_node(node))
return NULL;
if (mas->index > max)
return NULL;
if (mas->offset > count)
return NULL;
while (mas->offset < count) {
pivot = pivots[mas->offset];
entry = mas_slot(mas, slots, mas->offset);
if (ma_dead_node(node))
return NULL;
if (entry)
goto found;
if (pivot >= max)
return NULL;
mas->index = pivot + 1;
mas->offset++;
}
if (mas->index > mas->max) {
mas->index = mas->last;
return NULL;
}
pivot = mas_safe_pivot(mas, pivots, mas->offset, type);
entry = mas_slot(mas, slots, mas->offset);
if (ma_dead_node(node))
return NULL;
if (!pivot)
return NULL;
if (!entry)
return NULL;
found:
mas->last = pivot;
return entry;
}
static inline void mas_rewalk(struct ma_state *mas, unsigned long index)
{
retry:
mas_set(mas, index);
mas_state_walk(mas);
if (mas_is_start(mas))
goto retry;
}
/*
* mas_next_entry() - Internal function to get the next entry.
* @mas: The maple state
* @limit: The maximum range start.
*
* Set the @mas->node to the next entry and the range_start to
* the beginning value for the entry. Does not check beyond @limit.
* Sets @mas->index and @mas->last to the limit if it is hit.
* Restarts on dead nodes.
*
* Return: the next entry or %NULL.
*/
static inline void *mas_next_entry(struct ma_state *mas, unsigned long limit)
{
void *entry = NULL;
struct maple_enode *prev_node;
struct maple_node *node;
unsigned char offset;
unsigned long last;
enum maple_type mt;
if (mas->index > limit) {
mas->index = mas->last = limit;
mas_pause(mas);
return NULL;
}
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
last = mas->last;
retry:
offset = mas->offset;
prev_node = mas->node;
node = mas_mn(mas);
mt = mte_node_type(mas->node);
mas->offset++;
if (unlikely(mas->offset >= mt_slots[mt])) {
mas->offset = mt_slots[mt] - 1;
goto next_node;
}
while (!mas_is_none(mas)) {
entry = mas_next_nentry(mas, node, limit, mt);
if (unlikely(ma_dead_node(node))) {
mas_rewalk(mas, last);
goto retry;
}
if (likely(entry))
return entry;
if (unlikely((mas->index > limit)))
break;
next_node:
prev_node = mas->node;
offset = mas->offset;
if (unlikely(mas_next_node(mas, node, limit))) {
mas_rewalk(mas, last);
goto retry;
}
mas->offset = 0;
node = mas_mn(mas);
mt = mte_node_type(mas->node);
}
mas->index = mas->last = limit;
mas->offset = offset;
mas->node = prev_node;
return NULL;
}
/*
* mas_prev_nentry() - Get the previous node entry.
* @mas: The maple state.
* @limit: The lower limit to check for a value.
*
* Return: the entry, %NULL otherwise.
*/
static inline void *mas_prev_nentry(struct ma_state *mas, unsigned long limit,
unsigned long index)
{
unsigned long pivot, min;
unsigned char offset;
struct maple_node *mn;
enum maple_type mt;
unsigned long *pivots;
void __rcu **slots;
void *entry;
retry:
if (!mas->offset)
return NULL;
mn = mas_mn(mas);
mt = mte_node_type(mas->node);
offset = mas->offset - 1;
if (offset >= mt_slots[mt])
offset = mt_slots[mt] - 1;
slots = ma_slots(mn, mt);
pivots = ma_pivots(mn, mt);
if (offset == mt_pivots[mt])
pivot = mas->max;
else
pivot = pivots[offset];
if (unlikely(ma_dead_node(mn))) {
mas_rewalk(mas, index);
goto retry;
}
while (offset && ((!mas_slot(mas, slots, offset) && pivot >= limit) ||
!pivot))
pivot = pivots[--offset];
min = mas_safe_min(mas, pivots, offset);
entry = mas_slot(mas, slots, offset);
if (unlikely(ma_dead_node(mn))) {
mas_rewalk(mas, index);
goto retry;
}
if (likely(entry)) {
mas->offset = offset;
mas->last = pivot;
mas->index = min;
}
return entry;
}
static inline void *mas_prev_entry(struct ma_state *mas, unsigned long min)
{
void *entry;
if (mas->index < min) {
mas->index = mas->last = min;
mas->node = MAS_NONE;
return NULL;
}
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
retry:
while (likely(!mas_is_none(mas))) {
entry = mas_prev_nentry(mas, min, mas->index);
if (unlikely(mas->last < min))
goto not_found;
if (likely(entry))
return entry;
if (unlikely(mas_prev_node(mas, min))) {
mas_rewalk(mas, mas->index);
goto retry;
}
mas->offset++;
}
mas->offset--;
not_found:
mas->index = mas->last = min;
return NULL;
}
/*
* mas_rev_awalk() - Internal function. Reverse allocation walk. Find the
* highest gap address of a given size in a given node and descend.
* @mas: The maple state
* @size: The needed size.
*
* Return: True if found in a leaf, false otherwise.
*
*/
static bool mas_rev_awalk(struct ma_state *mas, unsigned long size)
{
enum maple_type type = mte_node_type(mas->node);
struct maple_node *node = mas_mn(mas);
unsigned long *pivots, *gaps;
void __rcu **slots;
unsigned long gap = 0;
unsigned long max, min;
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
unsigned char offset;
if (unlikely(mas_is_err(mas)))
return true;
if (ma_is_dense(type)) {
/* dense nodes. */
mas->offset = (unsigned char)(mas->index - mas->min);
return true;
}
pivots = ma_pivots(node, type);
slots = ma_slots(node, type);
gaps = ma_gaps(node, type);
offset = mas->offset;
min = mas_safe_min(mas, pivots, offset);
/* Skip out of bounds. */
while (mas->last < min)
min = mas_safe_min(mas, pivots, --offset);
max = mas_safe_pivot(mas, pivots, offset, type);
while (mas->index <= max) {
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
gap = 0;
if (gaps)
gap = gaps[offset];
else if (!mas_slot(mas, slots, offset))
gap = max - min + 1;
if (gap) {
if ((size <= gap) && (size <= mas->last - min + 1))
break;
if (!gaps) {
/* Skip the next slot, it cannot be a gap. */
if (offset < 2)
goto ascend;
offset -= 2;
max = pivots[offset];
min = mas_safe_min(mas, pivots, offset);
continue;
}
}
if (!offset)
goto ascend;
offset--;
max = min - 1;
min = mas_safe_min(mas, pivots, offset);
}
if (unlikely((mas->index > max) || (size - 1 > max - mas->index)))
goto no_space;
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
if (unlikely(ma_is_leaf(type))) {
mas->offset = offset;
mas->min = min;
mas->max = min + gap - 1;
return true;
}
/* descend, only happens under lock. */
mas->node = mas_slot(mas, slots, offset);
mas->min = min;
mas->max = max;
mas->offset = mas_data_end(mas);
return false;
ascend:
if (!mte_is_root(mas->node))
return false;
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
no_space:
mas_set_err(mas, -EBUSY);
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
return false;
}
static inline bool mas_anode_descend(struct ma_state *mas, unsigned long size)
{
enum maple_type type = mte_node_type(mas->node);
unsigned long pivot, min, gap = 0;
unsigned char offset;
unsigned long *gaps;
unsigned long *pivots = ma_pivots(mas_mn(mas), type);
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
void __rcu **slots = ma_slots(mas_mn(mas), type);
bool found = false;
if (ma_is_dense(type)) {
mas->offset = (unsigned char)(mas->index - mas->min);
return true;
}
gaps = ma_gaps(mte_to_node(mas->node), type);
offset = mas->offset;
min = mas_safe_min(mas, pivots, offset);
for (; offset < mt_slots[type]; offset++) {
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
pivot = mas_safe_pivot(mas, pivots, offset, type);
if (offset && !pivot)
break;
/* Not within lower bounds */
if (mas->index > pivot)
goto next_slot;
if (gaps)
gap = gaps[offset];
else if (!mas_slot(mas, slots, offset))
gap = min(pivot, mas->last) - max(mas->index, min) + 1;
else
goto next_slot;
if (gap >= size) {
if (ma_is_leaf(type)) {
found = true;
goto done;
}
if (mas->index <= pivot) {
mas->node = mas_slot(mas, slots, offset);
mas->min = min;
mas->max = pivot;
offset = 0;
break;
}
}
next_slot:
min = pivot + 1;
if (mas->last <= pivot) {
mas_set_err(mas, -EBUSY);
return true;
}
}
if (mte_is_root(mas->node))
found = true;
done:
mas->offset = offset;
return found;
}
/**
* mas_walk() - Search for @mas->index in the tree.
* @mas: The maple state.
*
* mas->index and mas->last will be set to the range if there is a value. If
* mas->node is MAS_NONE, reset to MAS_START.
*
* Return: the entry at the location or %NULL.
*/
void *mas_walk(struct ma_state *mas)
{
void *entry;
retry:
entry = mas_state_walk(mas);
if (mas_is_start(mas))
goto retry;
if (mas_is_ptr(mas)) {
if (!mas->index) {
mas->last = 0;
} else {
mas->index = 1;
mas->last = ULONG_MAX;
}
return entry;
}
if (mas_is_none(mas)) {
mas->index = 0;
mas->last = ULONG_MAX;
}
return entry;
}
EXPORT_SYMBOL_GPL(mas_walk);
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
static inline bool mas_rewind_node(struct ma_state *mas)
{
unsigned char slot;
do {
if (mte_is_root(mas->node)) {
slot = mas->offset;
if (!slot)
return false;
} else {
mas_ascend(mas);
slot = mas->offset;
}
} while (!slot);
mas->offset = --slot;
return true;
}
/*
* mas_skip_node() - Internal function. Skip over a node.
* @mas: The maple state.
*
* Return: true if there is another node, false otherwise.
*/
static inline bool mas_skip_node(struct ma_state *mas)
{
unsigned char slot, slot_count;
unsigned long *pivots;
enum maple_type mt;
mt = mte_node_type(mas->node);
slot_count = mt_slots[mt] - 1;
do {
if (mte_is_root(mas->node)) {
slot = mas->offset;
if (slot > slot_count) {
mas_set_err(mas, -EBUSY);
return false;
}
} else {
mas_ascend(mas);
slot = mas->offset;
mt = mte_node_type(mas->node);
slot_count = mt_slots[mt] - 1;
}
} while (slot > slot_count);
mas->offset = ++slot;
pivots = ma_pivots(mas_mn(mas), mt);
if (slot > 0)
mas->min = pivots[slot - 1] + 1;
if (slot <= slot_count)
mas->max = pivots[slot];
return true;
}
/*
* mas_awalk() - Allocation walk. Search from low address to high, for a gap of
* @size
* @mas: The maple state
* @size: The size of the gap required
*
* Search between @mas->index and @mas->last for a gap of @size.
*/
static inline void mas_awalk(struct ma_state *mas, unsigned long size)
{
struct maple_enode *last = NULL;
/*
* There are 4 options:
* go to child (descend)
* go back to parent (ascend)
* no gap found. (return, slot == MAPLE_NODE_SLOTS)
* found the gap. (return, slot != MAPLE_NODE_SLOTS)
*/
while (!mas_is_err(mas) && !mas_anode_descend(mas, size)) {
if (last == mas->node)
mas_skip_node(mas);
else
last = mas->node;
}
}
/*
* mas_fill_gap() - Fill a located gap with @entry.
* @mas: The maple state
* @entry: The value to store
* @slot: The offset into the node to store the @entry
* @size: The size of the entry
* @index: The start location
*/
static inline void mas_fill_gap(struct ma_state *mas, void *entry,
unsigned char slot, unsigned long size, unsigned long *index)
{
MA_WR_STATE(wr_mas, mas, entry);
unsigned char pslot = mte_parent_slot(mas->node);
struct maple_enode *mn = mas->node;
unsigned long *pivots;
enum maple_type ptype;
/*
* mas->index is the start address for the search
* which may no longer be needed.
* mas->last is the end address for the search
*/
*index = mas->index;
mas->last = mas->index + size - 1;
/*
* It is possible that using mas->max and mas->min to correctly
* calculate the index and last will cause an issue in the gap
* calculation, so fix the ma_state here
*/
mas_ascend(mas);
ptype = mte_node_type(mas->node);
pivots = ma_pivots(mas_mn(mas), ptype);
mas->max = mas_safe_pivot(mas, pivots, pslot, ptype);
mas->min = mas_safe_min(mas, pivots, pslot);
mas->node = mn;
mas->offset = slot;
mas_wr_store_entry(&wr_mas);
}
/*
* mas_sparse_area() - Internal function. Return upper or lower limit when
* searching for a gap in an empty tree.
* @mas: The maple state
* @min: the minimum range
* @max: The maximum range
* @size: The size of the gap
* @fwd: Searching forward or back
*/
static inline void mas_sparse_area(struct ma_state *mas, unsigned long min,
unsigned long max, unsigned long size, bool fwd)
{
unsigned long start = 0;
if (!unlikely(mas_is_none(mas)))
start++;
/* mas_is_ptr */
if (start < min)
start = min;
if (fwd) {
mas->index = start;
mas->last = start + size - 1;
return;
}
mas->index = max;
}
/*
* mas_empty_area() - Get the lowest address within the range that is
* sufficient for the size requested.
* @mas: The maple state
* @min: The lowest value of the range
* @max: The highest value of the range
* @size: The size needed
*/
int mas_empty_area(struct ma_state *mas, unsigned long min,
unsigned long max, unsigned long size)
{
unsigned char offset;
unsigned long *pivots;
enum maple_type mt;
if (mas_is_start(mas))
mas_start(mas);
else if (mas->offset >= 2)
mas->offset -= 2;
else if (!mas_skip_node(mas))
return -EBUSY;
/* Empty set */
if (mas_is_none(mas) || mas_is_ptr(mas)) {
mas_sparse_area(mas, min, max, size, true);
return 0;
}
/* The start of the window can only be within these values */
mas->index = min;
mas->last = max;
mas_awalk(mas, size);
if (unlikely(mas_is_err(mas)))
return xa_err(mas->node);
offset = mas->offset;
if (unlikely(offset == MAPLE_NODE_SLOTS))
return -EBUSY;
mt = mte_node_type(mas->node);
pivots = ma_pivots(mas_mn(mas), mt);
if (offset)
mas->min = pivots[offset - 1] + 1;
if (offset < mt_pivots[mt])
mas->max = pivots[offset];
if (mas->index < mas->min)
mas->index = mas->min;
mas->last = mas->index + size - 1;
return 0;
}
EXPORT_SYMBOL_GPL(mas_empty_area);
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
/*
* mas_empty_area_rev() - Get the highest address within the range that is
* sufficient for the size requested.
* @mas: The maple state
* @min: The lowest value of the range
* @max: The highest value of the range
* @size: The size needed
*/
int mas_empty_area_rev(struct ma_state *mas, unsigned long min,
unsigned long max, unsigned long size)
{
struct maple_enode *last = mas->node;
if (mas_is_start(mas)) {
mas_start(mas);
mas->offset = mas_data_end(mas);
} else if (mas->offset >= 2) {
mas->offset -= 2;
} else if (!mas_rewind_node(mas)) {
return -EBUSY;
}
/* Empty set. */
if (mas_is_none(mas) || mas_is_ptr(mas)) {
mas_sparse_area(mas, min, max, size, false);
return 0;
}
/* The start of the window can only be within these values. */
mas->index = min;
mas->last = max;
while (!mas_rev_awalk(mas, size)) {
if (last == mas->node) {
if (!mas_rewind_node(mas))
return -EBUSY;
} else {
last = mas->node;
}
}
if (mas_is_err(mas))
return xa_err(mas->node);
if (unlikely(mas->offset == MAPLE_NODE_SLOTS))
return -EBUSY;
/*
* mas_rev_awalk() has set mas->min and mas->max to the gap values. If
* the maximum is outside the window we are searching, then use the last
* location in the search.
* mas->max and mas->min is the range of the gap.
* mas->index and mas->last are currently set to the search range.
*/
/* Trim the upper limit to the max. */
if (mas->max <= mas->last)
mas->last = mas->max;
mas->index = mas->last - size + 1;
return 0;
}
EXPORT_SYMBOL_GPL(mas_empty_area_rev);
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
static inline int mas_alloc(struct ma_state *mas, void *entry,
unsigned long size, unsigned long *index)
{
unsigned long min;
mas_start(mas);
if (mas_is_none(mas) || mas_is_ptr(mas)) {
mas_root_expand(mas, entry);
if (mas_is_err(mas))
return xa_err(mas->node);
if (!mas->index)
return mte_pivot(mas->node, 0);
return mte_pivot(mas->node, 1);
}
/* Must be walking a tree. */
mas_awalk(mas, size);
if (mas_is_err(mas))
return xa_err(mas->node);
if (mas->offset == MAPLE_NODE_SLOTS)
goto no_gap;
/*
* At this point, mas->node points to the right node and we have an
* offset that has a sufficient gap.
*/
min = mas->min;
if (mas->offset)
min = mte_pivot(mas->node, mas->offset - 1) + 1;
if (mas->index < min)
mas->index = min;
mas_fill_gap(mas, entry, mas->offset, size, index);
return 0;
no_gap:
return -EBUSY;
}
static inline int mas_rev_alloc(struct ma_state *mas, unsigned long min,
unsigned long max, void *entry,
unsigned long size, unsigned long *index)
{
int ret = 0;
ret = mas_empty_area_rev(mas, min, max, size);
if (ret)
return ret;
if (mas_is_err(mas))
return xa_err(mas->node);
if (mas->offset == MAPLE_NODE_SLOTS)
goto no_gap;
mas_fill_gap(mas, entry, mas->offset, size, index);
return 0;
no_gap:
return -EBUSY;
}
/*
* mas_dead_leaves() - Mark all leaves of a node as dead.
* @mas: The maple state
* @slots: Pointer to the slot array
*
* Must hold the write lock.
*
* Return: The number of leaves marked as dead.
*/
static inline
unsigned char mas_dead_leaves(struct ma_state *mas, void __rcu **slots)
{
struct maple_node *node;
enum maple_type type;
void *entry;
int offset;
for (offset = 0; offset < mt_slot_count(mas->node); offset++) {
entry = mas_slot_locked(mas, slots, offset);
type = mte_node_type(entry);
node = mte_to_node(entry);
/* Use both node and type to catch LE & BE metadata */
if (!node || !type)
break;
mte_set_node_dead(entry);
smp_wmb(); /* Needed for RCU */
node->type = type;
rcu_assign_pointer(slots[offset], node);
}
return offset;
}
static void __rcu **mas_dead_walk(struct ma_state *mas, unsigned char offset)
{
struct maple_node *node, *next;
void __rcu **slots = NULL;
next = mas_mn(mas);
do {
mas->node = ma_enode_ptr(next);
node = mas_mn(mas);
slots = ma_slots(node, node->type);
next = mas_slot_locked(mas, slots, offset);
offset = 0;
} while (!ma_is_leaf(next->type));
return slots;
}
static void mt_free_walk(struct rcu_head *head)
{
void __rcu **slots;
struct maple_node *node, *start;
struct maple_tree mt;
unsigned char offset;
enum maple_type type;
MA_STATE(mas, &mt, 0, 0);
node = container_of(head, struct maple_node, rcu);
if (ma_is_leaf(node->type))
goto free_leaf;
mt_init_flags(&mt, node->ma_flags);
mas_lock(&mas);
start = node;
mas.node = mt_mk_node(node, node->type);
slots = mas_dead_walk(&mas, 0);
node = mas_mn(&mas);
do {
mt_free_bulk(node->slot_len, slots);
offset = node->parent_slot + 1;
mas.node = node->piv_parent;
if (mas_mn(&mas) == node)
goto start_slots_free;
type = mte_node_type(mas.node);
slots = ma_slots(mte_to_node(mas.node), type);
if ((offset < mt_slots[type]) && (slots[offset]))
slots = mas_dead_walk(&mas, offset);
node = mas_mn(&mas);
} while ((node != start) || (node->slot_len < offset));
slots = ma_slots(node, node->type);
mt_free_bulk(node->slot_len, slots);
start_slots_free:
mas_unlock(&mas);
free_leaf:
mt_free_rcu(&node->rcu);
}
static inline void __rcu **mas_destroy_descend(struct ma_state *mas,
struct maple_enode *prev, unsigned char offset)
{
struct maple_node *node;
struct maple_enode *next = mas->node;
void __rcu **slots = NULL;
do {
mas->node = next;
node = mas_mn(mas);
slots = ma_slots(node, mte_node_type(mas->node));
next = mas_slot_locked(mas, slots, 0);
if ((mte_dead_node(next)))
next = mas_slot_locked(mas, slots, 1);
mte_set_node_dead(mas->node);
node->type = mte_node_type(mas->node);
node->piv_parent = prev;
node->parent_slot = offset;
offset = 0;
prev = mas->node;
} while (!mte_is_leaf(next));
return slots;
}
static void mt_destroy_walk(struct maple_enode *enode, unsigned char ma_flags,
bool free)
{
void __rcu **slots;
struct maple_node *node = mte_to_node(enode);
struct maple_enode *start;
struct maple_tree mt;
MA_STATE(mas, &mt, 0, 0);
if (mte_is_leaf(enode))
goto free_leaf;
mt_init_flags(&mt, ma_flags);
mas_lock(&mas);
mas.node = start = enode;
slots = mas_destroy_descend(&mas, start, 0);
node = mas_mn(&mas);
do {
enum maple_type type;
unsigned char offset;
struct maple_enode *parent, *tmp;
node->slot_len = mas_dead_leaves(&mas, slots);
if (free)
mt_free_bulk(node->slot_len, slots);
offset = node->parent_slot + 1;
mas.node = node->piv_parent;
if (mas_mn(&mas) == node)
goto start_slots_free;
type = mte_node_type(mas.node);
slots = ma_slots(mte_to_node(mas.node), type);
if (offset >= mt_slots[type])
goto next;
tmp = mas_slot_locked(&mas, slots, offset);
if (mte_node_type(tmp) && mte_to_node(tmp)) {
parent = mas.node;
mas.node = tmp;
slots = mas_destroy_descend(&mas, parent, offset);
}
next:
node = mas_mn(&mas);
} while (start != mas.node);
node = mas_mn(&mas);
node->slot_len = mas_dead_leaves(&mas, slots);
if (free)
mt_free_bulk(node->slot_len, slots);
start_slots_free:
mas_unlock(&mas);
free_leaf:
if (free)
mt_free_rcu(&node->rcu);
}
/*
* mte_destroy_walk() - Free a tree or sub-tree.
* @enode: the encoded maple node (maple_enode) to start
* @mt: the tree to free - needed for node types.
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
*
* Must hold the write lock.
*/
static inline void mte_destroy_walk(struct maple_enode *enode,
struct maple_tree *mt)
{
struct maple_node *node = mte_to_node(enode);
if (mt_in_rcu(mt)) {
mt_destroy_walk(enode, mt->ma_flags, false);
call_rcu(&node->rcu, mt_free_walk);
} else {
mt_destroy_walk(enode, mt->ma_flags, true);
}
}
static void mas_wr_store_setup(struct ma_wr_state *wr_mas)
{
if (unlikely(mas_is_paused(wr_mas->mas)))
mas_reset(wr_mas->mas);
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
if (!mas_is_start(wr_mas->mas)) {
if (mas_is_none(wr_mas->mas)) {
mas_reset(wr_mas->mas);
} else {
wr_mas->r_max = wr_mas->mas->max;
wr_mas->type = mte_node_type(wr_mas->mas->node);
if (mas_is_span_wr(wr_mas))
mas_reset(wr_mas->mas);
}
}
}
/* Interface */
/**
* mas_store() - Store an @entry.
* @mas: The maple state.
* @entry: The entry to store.
*
* The @mas->index and @mas->last is used to set the range for the @entry.
* Note: The @mas should have pre-allocated entries to ensure there is memory to
* store the entry. Please see mas_expected_entries()/mas_destroy() for more details.
*
* Return: the first entry between mas->index and mas->last or %NULL.
*/
void *mas_store(struct ma_state *mas, void *entry)
{
MA_WR_STATE(wr_mas, mas, entry);
trace_ma_write(__func__, mas, 0, entry);
#ifdef CONFIG_DEBUG_MAPLE_TREE
if (mas->index > mas->last)
pr_err("Error %lu > %lu %p\n", mas->index, mas->last, entry);
MT_BUG_ON(mas->tree, mas->index > mas->last);
if (mas->index > mas->last) {
mas_set_err(mas, -EINVAL);
return NULL;
}
#endif
/*
* Storing is the same operation as insert with the added caveat that it
* can overwrite entries. Although this seems simple enough, one may
* want to examine what happens if a single store operation was to
* overwrite multiple entries within a self-balancing B-Tree.
*/
mas_wr_store_setup(&wr_mas);
mas_wr_store_entry(&wr_mas);
return wr_mas.content;
}
EXPORT_SYMBOL_GPL(mas_store);
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
/**
* mas_store_gfp() - Store a value into the tree.
* @mas: The maple state
* @entry: The entry to store
* @gfp: The GFP_FLAGS to use for allocations if necessary.
*
* Return: 0 on success, -EINVAL on invalid request, -ENOMEM if memory could not
* be allocated.
*/
int mas_store_gfp(struct ma_state *mas, void *entry, gfp_t gfp)
{
MA_WR_STATE(wr_mas, mas, entry);
mas_wr_store_setup(&wr_mas);
trace_ma_write(__func__, mas, 0, entry);
retry:
mas_wr_store_entry(&wr_mas);
if (unlikely(mas_nomem(mas, gfp)))
goto retry;
if (unlikely(mas_is_err(mas)))
return xa_err(mas->node);
return 0;
}
EXPORT_SYMBOL_GPL(mas_store_gfp);
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
/**
* mas_store_prealloc() - Store a value into the tree using memory
* preallocated in the maple state.
* @mas: The maple state
* @entry: The entry to store.
*/
void mas_store_prealloc(struct ma_state *mas, void *entry)
{
MA_WR_STATE(wr_mas, mas, entry);
mas_wr_store_setup(&wr_mas);
trace_ma_write(__func__, mas, 0, entry);
mas_wr_store_entry(&wr_mas);
BUG_ON(mas_is_err(mas));
mas_destroy(mas);
}
EXPORT_SYMBOL_GPL(mas_store_prealloc);
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
/**
* mas_preallocate() - Preallocate enough nodes for a store operation
* @mas: The maple state
* @gfp: The GFP_FLAGS to use for allocations.
*
* Return: 0 on success, -ENOMEM if memory could not be allocated.
*/
int mas_preallocate(struct ma_state *mas, gfp_t gfp)
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
{
int ret;
mas_node_count_gfp(mas, 1 + mas_mt_height(mas) * 3, gfp);
mas->mas_flags |= MA_STATE_PREALLOC;
if (likely(!mas_is_err(mas)))
return 0;
mas_set_alloc_req(mas, 0);
ret = xa_err(mas->node);
mas_reset(mas);
mas_destroy(mas);
mas_reset(mas);
return ret;
}
/*
* mas_destroy() - destroy a maple state.
* @mas: The maple state
*
* Upon completion, check the left-most node and rebalance against the node to
* the right if necessary. Frees any allocated nodes associated with this maple
* state.
*/
void mas_destroy(struct ma_state *mas)
{
struct maple_alloc *node;
unsigned long total;
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
/*
* When using mas_for_each() to insert an expected number of elements,
* it is possible that the number inserted is less than the expected
* number. To fix an invalid final node, a check is performed here to
* rebalance the previous node with the final node.
*/
if (mas->mas_flags & MA_STATE_REBALANCE) {
unsigned char end;
if (mas_is_start(mas))
mas_start(mas);
mtree_range_walk(mas);
end = mas_data_end(mas) + 1;
if (end < mt_min_slot_count(mas->node) - 1)
mas_destroy_rebalance(mas, end);
mas->mas_flags &= ~MA_STATE_REBALANCE;
}
mas->mas_flags &= ~(MA_STATE_BULK|MA_STATE_PREALLOC);
total = mas_allocated(mas);
while (total) {
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
node = mas->alloc;
mas->alloc = node->slot[0];
if (node->node_count > 1) {
size_t count = node->node_count - 1;
mt_free_bulk(count, (void __rcu **)&node->slot[1]);
total -= count;
}
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
kmem_cache_free(maple_node_cache, node);
total--;
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
}
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
mas->alloc = NULL;
}
EXPORT_SYMBOL_GPL(mas_destroy);
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
/*
* mas_expected_entries() - Set the expected number of entries that will be inserted.
* @mas: The maple state
* @nr_entries: The number of expected entries.
*
* This will attempt to pre-allocate enough nodes to store the expected number
* of entries. The allocations will occur using the bulk allocator interface
* for speed. Please call mas_destroy() on the @mas after inserting the entries
* to ensure any unused nodes are freed.
*
* Return: 0 on success, -ENOMEM if memory could not be allocated.
*/
int mas_expected_entries(struct ma_state *mas, unsigned long nr_entries)
{
int nonleaf_cap = MAPLE_ARANGE64_SLOTS - 2;
struct maple_enode *enode = mas->node;
int nr_nodes;
int ret;
/*
* Sometimes it is necessary to duplicate a tree to a new tree, such as
* forking a process and duplicating the VMAs from one tree to a new
* tree. When such a situation arises, it is known that the new tree is
* not going to be used until the entire tree is populated. For
* performance reasons, it is best to use a bulk load with RCU disabled.
* This allows for optimistic splitting that favours the left and reuse
* of nodes during the operation.
*/
/* Optimize splitting for bulk insert in-order */
mas->mas_flags |= MA_STATE_BULK;
/*
* Avoid overflow, assume a gap between each entry and a trailing null.
* If this is wrong, it just means allocation can happen during
* insertion of entries.
*/
nr_nodes = max(nr_entries, nr_entries * 2 + 1);
if (!mt_is_alloc(mas->tree))
nonleaf_cap = MAPLE_RANGE64_SLOTS - 2;
/* Leaves; reduce slots to keep space for expansion */
nr_nodes = DIV_ROUND_UP(nr_nodes, MAPLE_RANGE64_SLOTS - 2);
/* Internal nodes */
nr_nodes += DIV_ROUND_UP(nr_nodes, nonleaf_cap);
/* Add working room for split (2 nodes) + new parents */
mas_node_count(mas, nr_nodes + 3);
/* Detect if allocations run out */
mas->mas_flags |= MA_STATE_PREALLOC;
if (!mas_is_err(mas))
return 0;
ret = xa_err(mas->node);
mas->node = enode;
mas_destroy(mas);
return ret;
}
EXPORT_SYMBOL_GPL(mas_expected_entries);
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
/**
* mas_next() - Get the next entry.
* @mas: The maple state
* @max: The maximum index to check.
*
* Returns the next entry after @mas->index.
* Must hold rcu_read_lock or the write lock.
* Can return the zero entry.
*
* Return: The next entry or %NULL
*/
void *mas_next(struct ma_state *mas, unsigned long max)
{
if (mas_is_none(mas) || mas_is_paused(mas))
mas->node = MAS_START;
if (mas_is_start(mas))
mas_walk(mas); /* Retries on dead nodes handled by mas_walk */
if (mas_is_ptr(mas)) {
if (!mas->index) {
mas->index = 1;
mas->last = ULONG_MAX;
}
return NULL;
}
if (mas->last == ULONG_MAX)
return NULL;
/* Retries on dead nodes handled by mas_next_entry */
return mas_next_entry(mas, max);
}
EXPORT_SYMBOL_GPL(mas_next);
/**
* mt_next() - get the next value in the maple tree
* @mt: The maple tree
* @index: The start index
* @max: The maximum index to check
*
* Return: The entry at @index or higher, or %NULL if nothing is found.
*/
void *mt_next(struct maple_tree *mt, unsigned long index, unsigned long max)
{
void *entry = NULL;
MA_STATE(mas, mt, index, index);
rcu_read_lock();
entry = mas_next(&mas, max);
rcu_read_unlock();
return entry;
}
EXPORT_SYMBOL_GPL(mt_next);
/**
* mas_prev() - Get the previous entry
* @mas: The maple state
* @min: The minimum value to check.
*
* Must hold rcu_read_lock or the write lock.
* Will reset mas to MAS_START if the node is MAS_NONE. Will stop on not
* searchable nodes.
*
* Return: the previous value or %NULL.
*/
void *mas_prev(struct ma_state *mas, unsigned long min)
{
if (!mas->index) {
/* Nothing comes before 0 */
mas->last = 0;
mas->node = MAS_NONE;
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
return NULL;
}
if (unlikely(mas_is_ptr(mas)))
return NULL;
if (mas_is_none(mas) || mas_is_paused(mas))
mas->node = MAS_START;
if (mas_is_start(mas)) {
mas_walk(mas);
if (!mas->index)
return NULL;
}
if (mas_is_ptr(mas)) {
if (!mas->index) {
mas->last = 0;
return NULL;
}
mas->index = mas->last = 0;
return mas_root_locked(mas);
}
return mas_prev_entry(mas, min);
}
EXPORT_SYMBOL_GPL(mas_prev);
/**
* mt_prev() - get the previous value in the maple tree
* @mt: The maple tree
* @index: The start index
* @min: The minimum index to check
*
* Return: The entry at @index or lower, or %NULL if nothing is found.
*/
void *mt_prev(struct maple_tree *mt, unsigned long index, unsigned long min)
{
void *entry = NULL;
MA_STATE(mas, mt, index, index);
rcu_read_lock();
entry = mas_prev(&mas, min);
rcu_read_unlock();
return entry;
}
EXPORT_SYMBOL_GPL(mt_prev);
/**
* mas_pause() - Pause a mas_find/mas_for_each to drop the lock.
* @mas: The maple state to pause
*
* Some users need to pause a walk and drop the lock they're holding in
* order to yield to a higher priority thread or carry out an operation
* on an entry. Those users should call this function before they drop
* the lock. It resets the @mas to be suitable for the next iteration
* of the loop after the user has reacquired the lock. If most entries
* found during a walk require you to call mas_pause(), the mt_for_each()
* iterator may be more appropriate.
*
*/
void mas_pause(struct ma_state *mas)
{
mas->node = MAS_PAUSE;
}
EXPORT_SYMBOL_GPL(mas_pause);
/**
* mas_find() - On the first call, find the entry at or after mas->index up to
* %max. Otherwise, find the entry after mas->index.
* @mas: The maple state
* @max: The maximum value to check.
*
* Must hold rcu_read_lock or the write lock.
* If an entry exists, last and index are updated accordingly.
* May set @mas->node to MAS_NONE.
*
* Return: The entry or %NULL.
*/
void *mas_find(struct ma_state *mas, unsigned long max)
{
if (unlikely(mas_is_paused(mas))) {
if (unlikely(mas->last == ULONG_MAX)) {
mas->node = MAS_NONE;
return NULL;
}
mas->node = MAS_START;
mas->index = ++mas->last;
}
if (unlikely(mas_is_none(mas)))
mas->node = MAS_START;
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
if (unlikely(mas_is_start(mas))) {
/* First run or continue */
void *entry;
if (mas->index > max)
return NULL;
entry = mas_walk(mas);
if (entry)
return entry;
}
if (unlikely(!mas_searchable(mas)))
return NULL;
/* Retries on dead nodes handled by mas_next_entry */
return mas_next_entry(mas, max);
}
EXPORT_SYMBOL_GPL(mas_find);
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
/**
* mas_find_rev: On the first call, find the first non-null entry at or below
* mas->index down to %min. Otherwise find the first non-null entry below
* mas->index down to %min.
* @mas: The maple state
* @min: The minimum value to check.
*
* Must hold rcu_read_lock or the write lock.
* If an entry exists, last and index are updated accordingly.
* May set @mas->node to MAS_NONE.
*
* Return: The entry or %NULL.
*/
void *mas_find_rev(struct ma_state *mas, unsigned long min)
{
if (unlikely(mas_is_paused(mas))) {
if (unlikely(mas->last == ULONG_MAX)) {
mas->node = MAS_NONE;
return NULL;
}
mas->node = MAS_START;
mas->last = --mas->index;
}
if (unlikely(mas_is_start(mas))) {
/* First run or continue */
void *entry;
if (mas->index < min)
return NULL;
entry = mas_walk(mas);
if (entry)
return entry;
}
if (unlikely(!mas_searchable(mas)))
return NULL;
if (mas->index < min)
return NULL;
/* Retries on dead nodes handled by mas_prev_entry */
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
return mas_prev_entry(mas, min);
}
EXPORT_SYMBOL_GPL(mas_find_rev);
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
/**
* mas_erase() - Find the range in which index resides and erase the entire
* range.
* @mas: The maple state
*
* Must hold the write lock.
* Searches for @mas->index, sets @mas->index and @mas->last to the range and
* erases that range.
*
* Return: the entry that was erased or %NULL, @mas->index and @mas->last are updated.
*/
void *mas_erase(struct ma_state *mas)
{
void *entry;
MA_WR_STATE(wr_mas, mas, NULL);
if (mas_is_none(mas) || mas_is_paused(mas))
mas->node = MAS_START;
/* Retry unnecessary when holding the write lock. */
entry = mas_state_walk(mas);
if (!entry)
return NULL;
write_retry:
/* Must reset to ensure spanning writes of last slot are detected */
mas_reset(mas);
mas_wr_store_setup(&wr_mas);
mas_wr_store_entry(&wr_mas);
if (mas_nomem(mas, GFP_KERNEL))
goto write_retry;
return entry;
}
EXPORT_SYMBOL_GPL(mas_erase);
/**
* mas_nomem() - Check if there was an error allocating and do the allocation
* if necessary If there are allocations, then free them.
* @mas: The maple state
* @gfp: The GFP_FLAGS to use for allocations
* Return: true on allocation, false otherwise.
*/
bool mas_nomem(struct ma_state *mas, gfp_t gfp)
__must_hold(mas->tree->lock)
{
if (likely(mas->node != MA_ERROR(-ENOMEM))) {
mas_destroy(mas);
return false;
}
if (gfpflags_allow_blocking(gfp) && !mt_external_lock(mas->tree)) {
mtree_unlock(mas->tree);
mas_alloc_nodes(mas, gfp);
mtree_lock(mas->tree);
} else {
mas_alloc_nodes(mas, gfp);
}
if (!mas_allocated(mas))
return false;
mas->node = MAS_START;
return true;
}
void __init maple_tree_init(void)
{
maple_node_cache = kmem_cache_create("maple_node",
sizeof(struct maple_node), sizeof(struct maple_node),
SLAB_PANIC, NULL);
}
/**
* mtree_load() - Load a value stored in a maple tree
* @mt: The maple tree
* @index: The index to load
*
* Return: the entry or %NULL
*/
void *mtree_load(struct maple_tree *mt, unsigned long index)
{
MA_STATE(mas, mt, index, index);
void *entry;
trace_ma_read(__func__, &mas);
rcu_read_lock();
retry:
entry = mas_start(&mas);
if (unlikely(mas_is_none(&mas)))
goto unlock;
if (unlikely(mas_is_ptr(&mas))) {
if (index)
entry = NULL;
goto unlock;
}
entry = mtree_lookup_walk(&mas);
if (!entry && unlikely(mas_is_start(&mas)))
goto retry;
unlock:
rcu_read_unlock();
if (xa_is_zero(entry))
return NULL;
return entry;
}
EXPORT_SYMBOL(mtree_load);
/**
* mtree_store_range() - Store an entry at a given range.
* @mt: The maple tree
* @index: The start of the range
* @last: The end of the range
* @entry: The entry to store
* @gfp: The GFP_FLAGS to use for allocations
*
* Return: 0 on success, -EINVAL on invalid request, -ENOMEM if memory could not
* be allocated.
*/
int mtree_store_range(struct maple_tree *mt, unsigned long index,
unsigned long last, void *entry, gfp_t gfp)
{
MA_STATE(mas, mt, index, last);
MA_WR_STATE(wr_mas, &mas, entry);
trace_ma_write(__func__, &mas, 0, entry);
if (WARN_ON_ONCE(xa_is_advanced(entry)))
return -EINVAL;
if (index > last)
return -EINVAL;
mtree_lock(mt);
retry:
mas_wr_store_entry(&wr_mas);
if (mas_nomem(&mas, gfp))
goto retry;
mtree_unlock(mt);
if (mas_is_err(&mas))
return xa_err(mas.node);
return 0;
}
EXPORT_SYMBOL(mtree_store_range);
/**
* mtree_store() - Store an entry at a given index.
* @mt: The maple tree
* @index: The index to store the value
* @entry: The entry to store
* @gfp: The GFP_FLAGS to use for allocations
*
* Return: 0 on success, -EINVAL on invalid request, -ENOMEM if memory could not
* be allocated.
*/
int mtree_store(struct maple_tree *mt, unsigned long index, void *entry,
gfp_t gfp)
{
return mtree_store_range(mt, index, index, entry, gfp);
}
EXPORT_SYMBOL(mtree_store);
/**
* mtree_insert_range() - Insert an entry at a give range if there is no value.
* @mt: The maple tree
* @first: The start of the range
* @last: The end of the range
* @entry: The entry to store
* @gfp: The GFP_FLAGS to use for allocations.
*
* Return: 0 on success, -EEXISTS if the range is occupied, -EINVAL on invalid
* request, -ENOMEM if memory could not be allocated.
*/
int mtree_insert_range(struct maple_tree *mt, unsigned long first,
unsigned long last, void *entry, gfp_t gfp)
{
MA_STATE(ms, mt, first, last);
if (WARN_ON_ONCE(xa_is_advanced(entry)))
return -EINVAL;
if (first > last)
return -EINVAL;
mtree_lock(mt);
retry:
mas_insert(&ms, entry);
if (mas_nomem(&ms, gfp))
goto retry;
mtree_unlock(mt);
if (mas_is_err(&ms))
return xa_err(ms.node);
return 0;
}
EXPORT_SYMBOL(mtree_insert_range);
/**
* mtree_insert() - Insert an entry at a give index if there is no value.
* @mt: The maple tree
* @index : The index to store the value
* @entry: The entry to store
* @gfp: The FGP_FLAGS to use for allocations.
*
* Return: 0 on success, -EEXISTS if the range is occupied, -EINVAL on invalid
* request, -ENOMEM if memory could not be allocated.
*/
int mtree_insert(struct maple_tree *mt, unsigned long index, void *entry,
gfp_t gfp)
{
return mtree_insert_range(mt, index, index, entry, gfp);
}
EXPORT_SYMBOL(mtree_insert);
int mtree_alloc_range(struct maple_tree *mt, unsigned long *startp,
void *entry, unsigned long size, unsigned long min,
unsigned long max, gfp_t gfp)
{
int ret = 0;
MA_STATE(mas, mt, min, max - size);
if (!mt_is_alloc(mt))
return -EINVAL;
if (WARN_ON_ONCE(mt_is_reserved(entry)))
return -EINVAL;
if (min > max)
return -EINVAL;
if (max < size)
return -EINVAL;
if (!size)
return -EINVAL;
mtree_lock(mt);
retry:
mas.offset = 0;
mas.index = min;
mas.last = max - size;
ret = mas_alloc(&mas, entry, size, startp);
if (mas_nomem(&mas, gfp))
goto retry;
mtree_unlock(mt);
return ret;
}
EXPORT_SYMBOL(mtree_alloc_range);
int mtree_alloc_rrange(struct maple_tree *mt, unsigned long *startp,
void *entry, unsigned long size, unsigned long min,
unsigned long max, gfp_t gfp)
{
int ret = 0;
MA_STATE(mas, mt, min, max - size);
if (!mt_is_alloc(mt))
return -EINVAL;
if (WARN_ON_ONCE(mt_is_reserved(entry)))
return -EINVAL;
if (min >= max)
return -EINVAL;
if (max < size - 1)
return -EINVAL;
if (!size)
return -EINVAL;
mtree_lock(mt);
retry:
ret = mas_rev_alloc(&mas, min, max, entry, size, startp);
if (mas_nomem(&mas, gfp))
goto retry;
mtree_unlock(mt);
return ret;
}
EXPORT_SYMBOL(mtree_alloc_rrange);
/**
* mtree_erase() - Find an index and erase the entire range.
* @mt: The maple tree
* @index: The index to erase
*
* Erasing is the same as a walk to an entry then a store of a NULL to that
* ENTIRE range. In fact, it is implemented as such using the advanced API.
*
* Return: The entry stored at the @index or %NULL
*/
void *mtree_erase(struct maple_tree *mt, unsigned long index)
{
void *entry = NULL;
MA_STATE(mas, mt, index, index);
trace_ma_op(__func__, &mas);
mtree_lock(mt);
entry = mas_erase(&mas);
mtree_unlock(mt);
return entry;
}
EXPORT_SYMBOL(mtree_erase);
/**
* __mt_destroy() - Walk and free all nodes of a locked maple tree.
* @mt: The maple tree
*
* Note: Does not handle locking.
*/
void __mt_destroy(struct maple_tree *mt)
{
void *root = mt_root_locked(mt);
rcu_assign_pointer(mt->ma_root, NULL);
if (xa_is_node(root))
mte_destroy_walk(root, mt);
mt->ma_flags = 0;
}
EXPORT_SYMBOL_GPL(__mt_destroy);
/**
* mtree_destroy() - Destroy a maple tree
* @mt: The maple tree
*
* Frees all resources used by the tree. Handles locking.
*/
void mtree_destroy(struct maple_tree *mt)
{
mtree_lock(mt);
__mt_destroy(mt);
mtree_unlock(mt);
}
EXPORT_SYMBOL(mtree_destroy);
/**
* mt_find() - Search from the start up until an entry is found.
* @mt: The maple tree
* @index: Pointer which contains the start location of the search
* @max: The maximum value to check
*
* Handles locking. @index will be incremented to one beyond the range.
*
* Return: The entry at or after the @index or %NULL
*/
void *mt_find(struct maple_tree *mt, unsigned long *index, unsigned long max)
{
MA_STATE(mas, mt, *index, *index);
void *entry;
#ifdef CONFIG_DEBUG_MAPLE_TREE
unsigned long copy = *index;
#endif
trace_ma_read(__func__, &mas);
if ((*index) > max)
return NULL;
rcu_read_lock();
retry:
entry = mas_state_walk(&mas);
if (mas_is_start(&mas))
goto retry;
if (unlikely(xa_is_zero(entry)))
entry = NULL;
if (entry)
goto unlock;
while (mas_searchable(&mas) && (mas.index < max)) {
entry = mas_next_entry(&mas, max);
if (likely(entry && !xa_is_zero(entry)))
break;
}
if (unlikely(xa_is_zero(entry)))
entry = NULL;
unlock:
rcu_read_unlock();
if (likely(entry)) {
*index = mas.last + 1;
#ifdef CONFIG_DEBUG_MAPLE_TREE
if ((*index) && (*index) <= copy)
pr_err("index not increased! %lx <= %lx\n",
*index, copy);
MT_BUG_ON(mt, (*index) && ((*index) <= copy));
#endif
}
return entry;
}
EXPORT_SYMBOL(mt_find);
/**
* mt_find_after() - Search from the start up until an entry is found.
* @mt: The maple tree
* @index: Pointer which contains the start location of the search
* @max: The maximum value to check
*
* Handles locking, detects wrapping on index == 0
*
* Return: The entry at or after the @index or %NULL
*/
void *mt_find_after(struct maple_tree *mt, unsigned long *index,
unsigned long max)
{
if (!(*index))
return NULL;
return mt_find(mt, index, max);
}
EXPORT_SYMBOL(mt_find_after);
#ifdef CONFIG_DEBUG_MAPLE_TREE
atomic_t maple_tree_tests_run;
EXPORT_SYMBOL_GPL(maple_tree_tests_run);
atomic_t maple_tree_tests_passed;
EXPORT_SYMBOL_GPL(maple_tree_tests_passed);
#ifndef __KERNEL__
extern void kmem_cache_set_non_kernel(struct kmem_cache *, unsigned int);
void mt_set_non_kernel(unsigned int val)
{
kmem_cache_set_non_kernel(maple_node_cache, val);
}
extern unsigned long kmem_cache_get_alloc(struct kmem_cache *);
unsigned long mt_get_alloc_size(void)
{
return kmem_cache_get_alloc(maple_node_cache);
}
extern void kmem_cache_zero_nr_tallocated(struct kmem_cache *);
void mt_zero_nr_tallocated(void)
{
kmem_cache_zero_nr_tallocated(maple_node_cache);
}
extern unsigned int kmem_cache_nr_tallocated(struct kmem_cache *);
unsigned int mt_nr_tallocated(void)
{
return kmem_cache_nr_tallocated(maple_node_cache);
}
extern unsigned int kmem_cache_nr_allocated(struct kmem_cache *);
unsigned int mt_nr_allocated(void)
{
return kmem_cache_nr_allocated(maple_node_cache);
}
/*
* mas_dead_node() - Check if the maple state is pointing to a dead node.
* @mas: The maple state
* @index: The index to restore in @mas.
*
* Used in test code.
* Return: 1 if @mas has been reset to MAS_START, 0 otherwise.
*/
static inline int mas_dead_node(struct ma_state *mas, unsigned long index)
{
if (unlikely(!mas_searchable(mas) || mas_is_start(mas)))
return 0;
if (likely(!mte_dead_node(mas->node)))
return 0;
mas_rewalk(mas, index);
return 1;
}
void mt_cache_shrink(void)
{
}
#else
/*
* mt_cache_shrink() - For testing, don't use this.
*
* Certain testcases can trigger an OOM when combined with other memory
* debugging configuration options. This function is used to reduce the
* possibility of an out of memory even due to kmem_cache objects remaining
* around for longer than usual.
*/
void mt_cache_shrink(void)
{
kmem_cache_shrink(maple_node_cache);
}
EXPORT_SYMBOL_GPL(mt_cache_shrink);
#endif /* not defined __KERNEL__ */
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
/*
* mas_get_slot() - Get the entry in the maple state node stored at @offset.
* @mas: The maple state
* @offset: The offset into the slot array to fetch.
*
* Return: The entry stored at @offset.
*/
static inline struct maple_enode *mas_get_slot(struct ma_state *mas,
unsigned char offset)
{
return mas_slot(mas, ma_slots(mas_mn(mas), mte_node_type(mas->node)),
offset);
}
/*
* mas_first_entry() - Go the first leaf and find the first entry.
* @mas: the maple state.
* @limit: the maximum index to check.
* @*r_start: Pointer to set to the range start.
*
* Sets mas->offset to the offset of the entry, r_start to the range minimum.
*
* Return: The first entry or MAS_NONE.
*/
static inline void *mas_first_entry(struct ma_state *mas, struct maple_node *mn,
unsigned long limit, enum maple_type mt)
{
unsigned long max;
unsigned long *pivots;
void __rcu **slots;
void *entry = NULL;
mas->index = mas->min;
if (mas->index > limit)
goto none;
max = mas->max;
mas->offset = 0;
while (likely(!ma_is_leaf(mt))) {
MT_BUG_ON(mas->tree, mte_dead_node(mas->node));
slots = ma_slots(mn, mt);
pivots = ma_pivots(mn, mt);
max = pivots[0];
entry = mas_slot(mas, slots, 0);
if (unlikely(ma_dead_node(mn)))
return NULL;
mas->node = entry;
mn = mas_mn(mas);
mt = mte_node_type(mas->node);
}
MT_BUG_ON(mas->tree, mte_dead_node(mas->node));
mas->max = max;
slots = ma_slots(mn, mt);
entry = mas_slot(mas, slots, 0);
if (unlikely(ma_dead_node(mn)))
return NULL;
/* Slot 0 or 1 must be set */
if (mas->index > limit)
goto none;
if (likely(entry))
return entry;
pivots = ma_pivots(mn, mt);
mas->index = pivots[0] + 1;
mas->offset = 1;
entry = mas_slot(mas, slots, 1);
if (unlikely(ma_dead_node(mn)))
return NULL;
if (mas->index > limit)
goto none;
if (likely(entry))
return entry;
none:
if (likely(!ma_dead_node(mn)))
mas->node = MAS_NONE;
return NULL;
}
/* Depth first search, post-order */
static void mas_dfs_postorder(struct ma_state *mas, unsigned long max)
{
struct maple_enode *p = MAS_NONE, *mn = mas->node;
unsigned long p_min, p_max;
mas_next_node(mas, mas_mn(mas), max);
if (!mas_is_none(mas))
return;
if (mte_is_root(mn))
return;
mas->node = mn;
mas_ascend(mas);
while (mas->node != MAS_NONE) {
p = mas->node;
p_min = mas->min;
p_max = mas->max;
mas_prev_node(mas, 0);
}
if (p == MAS_NONE)
return;
mas->node = p;
mas->max = p_max;
mas->min = p_min;
}
/* Tree validations */
static void mt_dump_node(const struct maple_tree *mt, void *entry,
unsigned long min, unsigned long max, unsigned int depth);
static void mt_dump_range(unsigned long min, unsigned long max,
unsigned int depth)
{
static const char spaces[] = " ";
if (min == max)
pr_info("%.*s%lu: ", depth * 2, spaces, min);
else
pr_info("%.*s%lu-%lu: ", depth * 2, spaces, min, max);
}
static void mt_dump_entry(void *entry, unsigned long min, unsigned long max,
unsigned int depth)
{
mt_dump_range(min, max, depth);
if (xa_is_value(entry))
pr_cont("value %ld (0x%lx) [%p]\n", xa_to_value(entry),
xa_to_value(entry), entry);
else if (xa_is_zero(entry))
pr_cont("zero (%ld)\n", xa_to_internal(entry));
else if (mt_is_reserved(entry))
pr_cont("UNKNOWN ENTRY (%p)\n", entry);
else
pr_cont("%p\n", entry);
}
static void mt_dump_range64(const struct maple_tree *mt, void *entry,
unsigned long min, unsigned long max, unsigned int depth)
{
struct maple_range_64 *node = &mte_to_node(entry)->mr64;
bool leaf = mte_is_leaf(entry);
unsigned long first = min;
int i;
pr_cont(" contents: ");
for (i = 0; i < MAPLE_RANGE64_SLOTS - 1; i++)
pr_cont("%p %lu ", node->slot[i], node->pivot[i]);
pr_cont("%p\n", node->slot[i]);
for (i = 0; i < MAPLE_RANGE64_SLOTS; i++) {
unsigned long last = max;
if (i < (MAPLE_RANGE64_SLOTS - 1))
last = node->pivot[i];
else if (!node->slot[i] && max != mt_node_max(entry))
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
break;
if (last == 0 && i > 0)
break;
if (leaf)
mt_dump_entry(mt_slot(mt, node->slot, i),
first, last, depth + 1);
else if (node->slot[i])
mt_dump_node(mt, mt_slot(mt, node->slot, i),
first, last, depth + 1);
if (last == max)
break;
if (last > max) {
pr_err("node %p last (%lu) > max (%lu) at pivot %d!\n",
node, last, max, i);
break;
}
first = last + 1;
}
}
static void mt_dump_arange64(const struct maple_tree *mt, void *entry,
unsigned long min, unsigned long max, unsigned int depth)
{
struct maple_arange_64 *node = &mte_to_node(entry)->ma64;
bool leaf = mte_is_leaf(entry);
unsigned long first = min;
int i;
pr_cont(" contents: ");
for (i = 0; i < MAPLE_ARANGE64_SLOTS; i++)
pr_cont("%lu ", node->gap[i]);
pr_cont("| %02X %02X| ", node->meta.end, node->meta.gap);
for (i = 0; i < MAPLE_ARANGE64_SLOTS - 1; i++)
pr_cont("%p %lu ", node->slot[i], node->pivot[i]);
pr_cont("%p\n", node->slot[i]);
for (i = 0; i < MAPLE_ARANGE64_SLOTS; i++) {
unsigned long last = max;
if (i < (MAPLE_ARANGE64_SLOTS - 1))
last = node->pivot[i];
else if (!node->slot[i])
break;
if (last == 0 && i > 0)
break;
if (leaf)
mt_dump_entry(mt_slot(mt, node->slot, i),
first, last, depth + 1);
else if (node->slot[i])
mt_dump_node(mt, mt_slot(mt, node->slot, i),
first, last, depth + 1);
if (last == max)
break;
if (last > max) {
pr_err("node %p last (%lu) > max (%lu) at pivot %d!\n",
node, last, max, i);
break;
}
first = last + 1;
}
}
static void mt_dump_node(const struct maple_tree *mt, void *entry,
unsigned long min, unsigned long max, unsigned int depth)
{
struct maple_node *node = mte_to_node(entry);
unsigned int type = mte_node_type(entry);
unsigned int i;
mt_dump_range(min, max, depth);
pr_cont("node %p depth %d type %d parent %p", node, depth, type,
node ? node->parent : NULL);
switch (type) {
case maple_dense:
pr_cont("\n");
for (i = 0; i < MAPLE_NODE_SLOTS; i++) {
if (min + i > max)
pr_cont("OUT OF RANGE: ");
mt_dump_entry(mt_slot(mt, node->slot, i),
min + i, min + i, depth);
}
break;
case maple_leaf_64:
case maple_range_64:
mt_dump_range64(mt, entry, min, max, depth);
break;
case maple_arange_64:
mt_dump_arange64(mt, entry, min, max, depth);
break;
default:
pr_cont(" UNKNOWN TYPE\n");
}
}
void mt_dump(const struct maple_tree *mt)
{
void *entry = rcu_dereference_check(mt->ma_root, mt_locked(mt));
pr_info("maple_tree(%p) flags %X, height %u root %p\n",
mt, mt->ma_flags, mt_height(mt), entry);
if (!xa_is_node(entry))
mt_dump_entry(entry, 0, 0, 0);
else if (entry)
mt_dump_node(mt, entry, 0, mt_node_max(entry), 0);
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
}
EXPORT_SYMBOL_GPL(mt_dump);
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
/*
* Calculate the maximum gap in a node and check if that's what is reported in
* the parent (unless root).
*/
static void mas_validate_gaps(struct ma_state *mas)
{
struct maple_enode *mte = mas->node;
struct maple_node *p_mn;
unsigned long gap = 0, max_gap = 0;
unsigned long p_end, p_start = mas->min;
unsigned char p_slot;
unsigned long *gaps = NULL;
unsigned long *pivots = ma_pivots(mte_to_node(mte), mte_node_type(mte));
int i;
if (ma_is_dense(mte_node_type(mte))) {
for (i = 0; i < mt_slot_count(mte); i++) {
if (mas_get_slot(mas, i)) {
if (gap > max_gap)
max_gap = gap;
gap = 0;
continue;
}
gap++;
}
goto counted;
}
gaps = ma_gaps(mte_to_node(mte), mte_node_type(mte));
for (i = 0; i < mt_slot_count(mte); i++) {
p_end = mas_logical_pivot(mas, pivots, i, mte_node_type(mte));
if (!gaps) {
if (mas_get_slot(mas, i)) {
gap = 0;
goto not_empty;
}
gap += p_end - p_start + 1;
} else {
void *entry = mas_get_slot(mas, i);
gap = gaps[i];
if (!entry) {
if (gap != p_end - p_start + 1) {
pr_err("%p[%u] -> %p %lu != %lu - %lu + 1\n",
mas_mn(mas), i,
mas_get_slot(mas, i), gap,
p_end, p_start);
mt_dump(mas->tree);
MT_BUG_ON(mas->tree,
gap != p_end - p_start + 1);
}
} else {
if (gap > p_end - p_start + 1) {
pr_err("%p[%u] %lu >= %lu - %lu + 1 (%lu)\n",
mas_mn(mas), i, gap, p_end, p_start,
p_end - p_start + 1);
MT_BUG_ON(mas->tree,
gap > p_end - p_start + 1);
}
}
}
if (gap > max_gap)
max_gap = gap;
not_empty:
p_start = p_end + 1;
if (p_end >= mas->max)
break;
}
counted:
if (mte_is_root(mte))
return;
p_slot = mte_parent_slot(mas->node);
p_mn = mte_parent(mte);
MT_BUG_ON(mas->tree, max_gap > mas->max);
if (ma_gaps(p_mn, mas_parent_enum(mas, mte))[p_slot] != max_gap) {
pr_err("gap %p[%u] != %lu\n", p_mn, p_slot, max_gap);
mt_dump(mas->tree);
}
MT_BUG_ON(mas->tree,
ma_gaps(p_mn, mas_parent_enum(mas, mte))[p_slot] != max_gap);
}
static void mas_validate_parent_slot(struct ma_state *mas)
{
struct maple_node *parent;
struct maple_enode *node;
enum maple_type p_type = mas_parent_enum(mas, mas->node);
unsigned char p_slot = mte_parent_slot(mas->node);
void __rcu **slots;
int i;
if (mte_is_root(mas->node))
return;
parent = mte_parent(mas->node);
slots = ma_slots(parent, p_type);
MT_BUG_ON(mas->tree, mas_mn(mas) == parent);
/* Check prev/next parent slot for duplicate node entry */
for (i = 0; i < mt_slots[p_type]; i++) {
node = mas_slot(mas, slots, i);
if (i == p_slot) {
if (node != mas->node)
pr_err("parent %p[%u] does not have %p\n",
parent, i, mas_mn(mas));
MT_BUG_ON(mas->tree, node != mas->node);
} else if (node == mas->node) {
pr_err("Invalid child %p at parent %p[%u] p_slot %u\n",
mas_mn(mas), parent, i, p_slot);
MT_BUG_ON(mas->tree, node == mas->node);
}
}
}
static void mas_validate_child_slot(struct ma_state *mas)
{
enum maple_type type = mte_node_type(mas->node);
void __rcu **slots = ma_slots(mte_to_node(mas->node), type);
unsigned long *pivots = ma_pivots(mte_to_node(mas->node), type);
struct maple_enode *child;
unsigned char i;
if (mte_is_leaf(mas->node))
return;
for (i = 0; i < mt_slots[type]; i++) {
child = mas_slot(mas, slots, i);
if (!pivots[i] || pivots[i] == mas->max)
break;
if (!child)
break;
if (mte_parent_slot(child) != i) {
pr_err("Slot error at %p[%u]: child %p has pslot %u\n",
mas_mn(mas), i, mte_to_node(child),
mte_parent_slot(child));
MT_BUG_ON(mas->tree, 1);
}
if (mte_parent(child) != mte_to_node(mas->node)) {
pr_err("child %p has parent %p not %p\n",
mte_to_node(child), mte_parent(child),
mte_to_node(mas->node));
MT_BUG_ON(mas->tree, 1);
}
}
}
/*
* Validate all pivots are within mas->min and mas->max.
*/
static void mas_validate_limits(struct ma_state *mas)
{
int i;
unsigned long prev_piv = 0;
enum maple_type type = mte_node_type(mas->node);
void __rcu **slots = ma_slots(mte_to_node(mas->node), type);
unsigned long *pivots = ma_pivots(mas_mn(mas), type);
/* all limits are fine here. */
if (mte_is_root(mas->node))
return;
for (i = 0; i < mt_slots[type]; i++) {
unsigned long piv;
piv = mas_safe_pivot(mas, pivots, i, type);
if (!piv && (i != 0))
break;
if (!mte_is_leaf(mas->node)) {
void *entry = mas_slot(mas, slots, i);
if (!entry)
pr_err("%p[%u] cannot be null\n",
mas_mn(mas), i);
MT_BUG_ON(mas->tree, !entry);
}
if (prev_piv > piv) {
pr_err("%p[%u] piv %lu < prev_piv %lu\n",
mas_mn(mas), i, piv, prev_piv);
MT_BUG_ON(mas->tree, piv < prev_piv);
}
if (piv < mas->min) {
pr_err("%p[%u] %lu < %lu\n", mas_mn(mas), i,
piv, mas->min);
MT_BUG_ON(mas->tree, piv < mas->min);
}
if (piv > mas->max) {
pr_err("%p[%u] %lu > %lu\n", mas_mn(mas), i,
piv, mas->max);
MT_BUG_ON(mas->tree, piv > mas->max);
}
prev_piv = piv;
if (piv == mas->max)
break;
}
for (i += 1; i < mt_slots[type]; i++) {
void *entry = mas_slot(mas, slots, i);
if (entry && (i != mt_slots[type] - 1)) {
pr_err("%p[%u] should not have entry %p\n", mas_mn(mas),
i, entry);
MT_BUG_ON(mas->tree, entry != NULL);
}
if (i < mt_pivots[type]) {
unsigned long piv = pivots[i];
if (!piv)
continue;
pr_err("%p[%u] should not have piv %lu\n",
mas_mn(mas), i, piv);
MT_BUG_ON(mas->tree, i < mt_pivots[type] - 1);
}
}
}
static void mt_validate_nulls(struct maple_tree *mt)
{
void *entry, *last = (void *)1;
unsigned char offset = 0;
void __rcu **slots;
MA_STATE(mas, mt, 0, 0);
mas_start(&mas);
if (mas_is_none(&mas) || (mas.node == MAS_ROOT))
return;
while (!mte_is_leaf(mas.node))
mas_descend(&mas);
slots = ma_slots(mte_to_node(mas.node), mte_node_type(mas.node));
do {
entry = mas_slot(&mas, slots, offset);
if (!last && !entry) {
pr_err("Sequential nulls end at %p[%u]\n",
mas_mn(&mas), offset);
}
MT_BUG_ON(mt, !last && !entry);
last = entry;
if (offset == mas_data_end(&mas)) {
mas_next_node(&mas, mas_mn(&mas), ULONG_MAX);
if (mas_is_none(&mas))
return;
offset = 0;
slots = ma_slots(mte_to_node(mas.node),
mte_node_type(mas.node));
} else {
offset++;
}
} while (!mas_is_none(&mas));
}
/*
* validate a maple tree by checking:
* 1. The limits (pivots are within mas->min to mas->max)
* 2. The gap is correctly set in the parents
*/
void mt_validate(struct maple_tree *mt)
{
unsigned char end;
MA_STATE(mas, mt, 0, 0);
rcu_read_lock();
mas_start(&mas);
if (!mas_searchable(&mas))
goto done;
mas_first_entry(&mas, mas_mn(&mas), ULONG_MAX, mte_node_type(mas.node));
while (!mas_is_none(&mas)) {
MT_BUG_ON(mas.tree, mte_dead_node(mas.node));
if (!mte_is_root(mas.node)) {
end = mas_data_end(&mas);
if ((end < mt_min_slot_count(mas.node)) &&
(mas.max != ULONG_MAX)) {
pr_err("Invalid size %u of %p\n", end,
mas_mn(&mas));
MT_BUG_ON(mas.tree, 1);
}
}
mas_validate_parent_slot(&mas);
mas_validate_child_slot(&mas);
mas_validate_limits(&mas);
if (mt_is_alloc(mt))
mas_validate_gaps(&mas);
mas_dfs_postorder(&mas, ULONG_MAX);
}
mt_validate_nulls(mt);
done:
rcu_read_unlock();
}
EXPORT_SYMBOL_GPL(mt_validate);
Maple Tree: add new data structure Patch series "Introducing the Maple Tree" The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. Davidlor said : Yes I like the maple tree, and at this stage I don't think we can ask for : more from this series wrt the MM - albeit there seems to still be some : folks reporting breakage. Fundamentally I see Liam's work to (re)move : complexity out of the MM (not to say that the actual maple tree is not : complex) by consolidating the three complimentary data structures very : much worth it considering performance does not take a hit. This was very : much a turn off with the range locking approach, which worst case scenario : incurred in prohibitive overhead. Also as Liam and Matthew have : mentioned, RCU opens up a lot of nice performance opportunities, and in : addition academia[1] has shown outstanding scalability of address spaces : with the foundation of replacing the locked rbtree with RCU aware trees. A similar work has been discovered in the academic press https://pdos.csail.mit.edu/papers/rcuvm:asplos12.pdf Sheer coincidence. We designed our tree with the intention of solving the hardest problem first. Upon settling on a b-tree variant and a rough outline, we researched ranged based b-trees and RCU b-trees and did find that article. So it was nice to find reassurances that we were on the right path, but our design choice of using ranges made that paper unusable for us. This patch (of 70): The maple tree is an RCU-safe range based B-tree designed to use modern processor cache efficiently. There are a number of places in the kernel that a non-overlapping range-based tree would be beneficial, especially one with a simple interface. If you use an rbtree with other data structures to improve performance or an interval tree to track non-overlapping ranges, then this is for you. The tree has a branching factor of 10 for non-leaf nodes and 16 for leaf nodes. With the increased branching factor, it is significantly shorter than the rbtree so it has fewer cache misses. The removal of the linked list between subsequent entries also reduces the cache misses and the need to pull in the previous and next VMA during many tree alterations. The first user that is covered in this patch set is the vm_area_struct, where three data structures are replaced by the maple tree: the augmented rbtree, the vma cache, and the linked list of VMAs in the mm_struct. The long term goal is to reduce or remove the mmap_lock contention. The plan is to get to the point where we use the maple tree in RCU mode. Readers will not block for writers. A single write operation will be allowed at a time. A reader re-walks if stale data is encountered. VMAs would be RCU enabled and this mode would be entered once multiple tasks are using the mm_struct. There is additional BUG_ON() calls added within the tree, most of which are in debug code. These will be replaced with a WARN_ON() call in the future. There is also additional BUG_ON() calls within the code which will also be reduced in number at a later date. These exist to catch things such as out-of-range accesses which would crash anyways. Link: https://lkml.kernel.org/r/20220906194824.2110408-1-Liam.Howlett@oracle.com Link: https://lkml.kernel.org/r/20220906194824.2110408-2-Liam.Howlett@oracle.com Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Sven Schnelle <svens@linux.ibm.com> Tested-by: Yu Zhao <yuzhao@google.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Hildenbrand <david@redhat.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: SeongJae Park <sj@kernel.org> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-07 03:48:39 +08:00
#endif /* CONFIG_DEBUG_MAPLE_TREE */